# Good in Progress > Good in Progress (goodinprogress.org) is a growing collection of true stories about ordinary people using AI to change what they care about: AI for good, told through the person, never the technology. Every story is told in the subject's own words, checked against primary sources before publication, and published with its sources listed at the bottom. _Last generated: 2026-10-11. This is the index generation date, not a new source-verification date._ Good in Progress is the counter-narrative to doom-and-hype AI coverage, earned with evidence rather than enthusiasm. Every story is about a person who saw a problem, decided not to wait for permission or a perfect system, and used the tools available to them to do something about it. The person is always the hero. AI is the tool. Editorial standard: every story starts with the subject's own words and is checked against other sources with no stake in how it turns out. Every claim traces to a direct quote from a listed source rather than a paraphrase, and each candidate passes a graded scoring rubric and a reputation check before it runs. Sources are listed on every story page, labeled where a source is the subject's funder, employer, or vendor, so a reader can weigh them. If a claim cannot be sourced, it does not run. Stories are filed under one of six lenses: Ordinary People (people), Social Impact (impact), Breakthrough (breakthrough), Community (community), Adaptation (adaptation), and Org Spotlight (org-uplift). This file (llms-full.txt) contains the full text of every story currently on the site, generated automatically from the site's own public content. See /llms.txt for the short index version. Read the collection with an ordinary HTTP GET. For Markdown at the same page URL, send Accept: text/markdown. HTML remains the default. GET and HEAD require no credentials. Missing document URLs return HTTP 404; a Markdown request gets recovery links to the index and sitemap. Cite the canonical story URL and its listed sources. Preserve source attribution, quotations, limitations, and the Sources last verified date when present. A page generation date does not mean the evidence was checked again. Do not turn a local program outcome into a company-wide claim. If a story says it was withdrawn, do not reuse its earlier claims. This site is an editorial collection, not an AI vendor, live news feed, or professional advice service. ## When to use Good in Progress - [Find practical examples](https://goodinprogress.org/stories): Look for people and communities using AI to address a concrete problem, for lessons, workshops, or evidence-based discussion. - [Explore organizational change](https://goodinprogress.org/stories?lens=org-uplift): Find published Org Spotlight stories about work and organizational choices. Check the scope and sources of each claim before generalizing. - [Check a story before citing it](https://goodinprogress.org/llms.txt): Find the canonical story URL, read the full card, and follow its source links to verify the claim you need. - [Understand the editorial approach](https://goodinprogress.org/about): Read the mission, six lenses, and verification method. - [Follow the research skill](https://goodinprogress.org/.well-known/agent-skills/find-gip-stories/SKILL.md): Find relevant stories, check their sources, and cite their outcomes with the right scope. - [Suggest a story](https://goodinprogress.org/nominate): Use the human-facing nomination form with the user's approval. Reading or fetching a page never submits a nomination; there is no public agent publishing interface. ## Collections - [How teachers and school teams are using AI](https://goodinprogress.org/topics/education): Meet teachers and school teams using AI for writing practice, chemistry songs, and tools they once bought from vendors. Read their stories and sources. - [Communities keeping their languages alive with AI](https://goodinprogress.org/topics/language-preservation): Read how Te Hiku Media, Michael Running Wolf, and Jared Coleman are building language tools with community knowledge, ownership, and control at the center. - [People building more accessible lives with AI](https://goodinprogress.org/topics/accessibility): Meet people using AI to find their way through shops and subways, describe images, read distant signs, and communicate. Explore their stories and sources. ## Stories ## The radio station that decided te reo Māori was worth building for URL: https://goodinprogress.org/stories/te-hiku-media-maori-ai Location: Kaitāia, Northland, New Zealand People: Peter-Lucas Jones Date: July 13, 2026 Dek: When global speech AI had no room for their language, a small radio station in Kaitāia built its own, and made sure the data stayed in Māori hands. - **The person and the place:** Peter-Lucas Jones runs Te Hiku Media, a Māori-run radio and TV station in Kaitāia, a small town at the top of New Zealand's North Island. For decades, the station held boxes of cassette tapes full of te reo Māori recordings that almost no one outside the community would ever hear. - **The problem Peter-Lucas cared about:** Te reo Māori is a language at risk, and the big speech AI models coming out of Silicon Valley weren't built with it in mind. Keoni Mahelona, the station's chief technology officer, has described what the elders decided about that archive: that the stories should be made accessible online for their own community. - **The agency moment:** Jones and Te Hiku Media didn't lobby Google or wait for Meta to notice. They decided to build their own speech model and to own the data behind it. As Jones put it, indigenous communities like his "don't only want to be the users of technology, we want to be the developers of it too." Then they called on their own community to help build it. - **What changed:** The community answered fast. By NVIDIA's account of the project, 2,500 people came forward, and inside ten days their readings from a 200,000-phrase bank added up to more than 300 hours of labeled speech. The model built on those voices gets te reo Māori transcription right 92% of the time, and NVIDIA puts bilingual English-te reo speech at 82%. Other indigenous communities took note: NVIDIA reports that Native Hawaiians and the Mohawk have speech projects of their own now underway. - **The move worth copying:** Jones has a plain way of explaining why ownership mattered as much as accuracy: "When we think about the ownership of data, when we think about intellectual property rights, data is like land." Nobody else was going to protect it for them, so they built it themselves. Jones's line is the one to carry: data is like land. If someone is deciding what happens to your community's, this is the version worth pointing at. - **Takeaway line:** "We want to be the developers of it too." — Peter-Lucas Jones, chief executive, Te Hiku Media, to RNZ Sources: - NVIDIA Blog, the company whose AI tools the project runs on: https://blogs.nvidia.com/blog/te-hiku-media-maori-speech-ai/ - RNZ: https://www.rnz.co.nz/news/te-ao-maori/527405/maori-media-company-head-peter-lucas-jones-named-on-time-magazine-list-for-preserving-te-reo-through-ai - NZ Herald / Northern Advocate: https://www.nzherald.co.nz/northern-advocate/news/te-hiku-media-ceo-peter-lucas-jones-on-time-magazines-time100-ai-list/JDG3MXQY5FGTJMBGS7PB4TBZ3Q/ Sources last verified: September 6, 2026 --- ## Kingsley Jasi didn't wait for an extension officer. He sent a photo instead. URL: https://goodinprogress.org/stories/kingsley-jasi-ulangizi-malawi Location: Chiradzulu, Malawi People: Kingsley Jasi Date: July 12, 2026 Dek: With roughly one agricultural extension officer for every 3,000 farmers in Malawi, a WhatsApp chatbot in Chichewa answered the question no one was coming to answer. - **The person and the place:** Kingsley Jasi raises corn and beans on his two-acre farm in Chiradzulu, Malawi. - **The problem Kingsley cared about:** Worms were eating his corn. He reached for his usual insecticide and it made things worse. Malawi has roughly one agricultural extension officer for every 3,000 farmers, by the Ministry of Agriculture's own count, so expert help wasn't coming anytime soon. - **The agency moment:** Instead of guessing again, Jasi took a photo of the damage and messaged it to Ulangizi, a WhatsApp chatbot that answers farming questions in Chichewa. The tool was built by the nonprofit Opportunity International on government guidelines and traditional know-how. - **What changed:** Ulangizi recommended a different insecticide, and it worked. "After following its recommendation, the worms were completely eliminated," Jasi said. "Since then, I have relied solely on the chatbot's guidance." The tool is honest about where it falls short, too: reporting on Ulangizi found it "can sometimes generate false answers and get confused by new questions," and it isn't always right. That kind of candor is rare in a farming tool. - **The move worth copying:** He didn't wait for someone to drive out with a manual. He took a photo and asked. Photograph the problem and ask. He had no specialist, just a phone and a plain question, and it was enough to save the crop. - **Takeaway line:** "After following its recommendation, the worms were completely eliminated." — Kingsley Jasi, farmer, Chiradzulu, Malawi, to Rest of World Sources: - Rest of World: https://restofworld.org/2025/malawi-ulangizi-ai-farming-chatbot/ - D+C Development and Cooperation: https://www.dandc.eu/en/article/new-ai-powered-app-ulangizi-helps-malawian-smallholder-farmers-get-real-time-advice-their - TIME, which credits its farmer quote to Bloomberg: https://time.com/7160849/opportunity-international-ai-farmers-teachers/ Sources last verified: August 13, 2026 --- ## David Hughes put the expert in the phone farmers already had URL: https://goodinprogress.org/stories/david-hughes-plantvillage-nuru Location: Kenya / Tanzania (built at Penn State, USA) People: David Hughes Date: July 12, 2026 Dek: When cassava disease outran the number of trained experts who could diagnose it, a Penn State professor decided the smartphone already in farmers' hands was the delivery system. - **The person and the place:** David Hughes is an entomology and biology professor at Penn State, working with cassava farmers in Kenya and Tanzania alongside the International Institute of Tropical Agriculture. - **The problem David cared about:** Cassava virus diseases develop slowly and are hard to spot, and by the time an untrained eye catches one, the crop is often already lost. Smallholder farmers face this largely on their own. Hughes described the gap plainly: they are farmers "who typically don't have access to the kind of resources, such as land-grant agricultural scientists and extension specialists, that American farmers have. But they do have smartphones in increasing numbers." - **The agency moment:** Hughes didn't wait for enough trained experts to exist. He looked at what farmers already carried, the smartphone, and built Nuru, a computer-vision diagnostic tool that runs fully offline on an ordinary Android phone. - **What changed:** In a peer-reviewed field study, the 2020 version of Nuru read a single cassava leaf correctly 65% of the time, against 40 to 58% for extension agents and 18 to 31% for farmers. The same paper is careful about the limits. Single-leaf diagnosis is unreliable, and on single leaves the app caught cassava brown streak disease only 21% of the time. Show it six leaves instead of one and accuracy climbs to 74 to 88%. Overall, the paper puts Nuru level with trained agricultural extension officers rather than ahead of them. The results have faces too: Reuters found cassava farmer Roseline Akochi, who says the app changed her harvests enough to pay school fees and build a house. - **The move worth copying:** Hughes didn't wait for enough experts to exist. He put the expert in the phone farmers already had, and said the goal plainly: "In the last 160 years, we've largely solved the food security problem in the United States, and now we want to be a 21st century land-grant institution." Put the expertise where the person already is, instead of waiting for enough experts to exist. If your field is short of specialists, that is the version worth pointing at. - **Takeaway line:** "I used to harvest 5 bags, now I harvest 200." — Roseline Akochi, cassava farmer, to Reuters Sources: - Penn State News, the researcher's own university newsroom: https://www.psu.edu/news/research/story/penn-state-developed-plant-disease-app-recognized-google - IITA, the project partner that co-developed the tool: https://www.iita.org/news-item/african-farmers-get-new-help-against-cassava-diseases-nuru-their-artificially-intelligent-assistant/ - Frontiers in Plant Science, the team's own paper: https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2020.590889/full - Reuters, via NTV Kenya: https://ntvkenya.co.ke/news/kenyan-farmers-embrace-ai-to-tackle-pests-and-crop-diseases/ - Brookings Foresight Africa commentary, by the Director General of IITA, the partner that co-developed Nuru: https://www.brookings.edu/articles/digital-solutions-in-agriculture-drive-meaningful-livelihood-improvements-for-african-smallholder-farmers/ - Mongabay: https://news.mongabay.com/2024/06/kenyan-farmers-turn-to-whatsapp-ai-tools-to-combat-crop-diseases/ Sources last verified: August 5, 2026 --- ## The 8,500 IKEA workers who got bigger jobs instead of redundancy notices URL: https://goodinprogress.org/stories/ikea-billie-reskilling Location: Global (Ingka Group; program 2021-2023) Date: July 12, 2026 Dek: When its chatbot Billie took over routine questions, IKEA's parent company retrained an entire call center function rather than cut it. - **The person and the place:** 8,500 call-center workers across Ingka Group, the retail company behind IKEA, between 2021 and 2023. Brandon Mikula was one of them. So was Richard Richardson, who works at a remote-sales center in Sheffield, England. - **The problem the call-center workers cared about:** Call center staff spent their days answering the same routine questions on repeat, work that wears people down without using what they actually know how to do. - **The agency moment:** When Billie, IKEA's chatbot, began resolving roughly half of those routine inquiries, Ingka didn't shrink the team. It retrained the group into remote interior design advisers, running design consultations and video calls, the complex problem-solving a chatbot can't do. Ulrika Biesèrt, the company's People and Culture Manager, set out the intent in Ingka's own announcement of the program. - **What changed:** From 2021 to 2023, Billie resolved roughly 47% of the customer enquiries it received, about 3.2 million of them, for nearly €13 million in reported savings. The workers went two ways. Some moved onto a resolutions team, handling the tangled cases Billie cannot close. Others became remote design advisers, running consultations over video. Three years on, Fortune went to a remote-sales centre and found them still doing it. Richard Richardson, at the Sheffield centre, told Fortune the AI is great at the by-the-book terms and conditions, but cannot yet work out what the customer has already been offered, or what else might resolve the problem. These are Ingka's own numbers, from one specific three-year program, not an audit of the company's employment record overall. This is the story of what happened to one group of workers, not a claim that AI has never cost a job at IKEA. - **The move worth copying:** Reskilling instead of layoffs was a choice Ingka made, and three years on, people from that group are still doing the work it trained them for. If your workplace is having that argument right now, this is the version worth pointing at. Know a company that made the same call? Feed the scout at goodinprogress.org. - **Takeaway line:** "We're committed to strengthening co-workers' employability in Ingka or elsewhere through lifelong learning and development and reskilling, and to accelerate the creation of new jobs." — Ulrika Biesèrt, People and Culture Manager, Ingka Group newsroom Sources: - Ingka Group newsroom, the company's own announcement: https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/ - Reuters, Helen Reid, 13 June 2023: https://www.reuters.com/technology/ikea-bets-remote-interior-design-ai-changes-sales-strategy-2023-06-13/ - Fortune, Claire Zillman, 30 July 2026: https://fortune.com/2026/07/30/ikea-ai-workforce-reskilling-jobs-billie-chatbot-global-500/ Sources last verified: August 7, 2026 --- ## Racing to read the rainforest's library before it burns URL: https://goodinprogress.org/stories/panama-peru-ai-moth-monitoring Location: Barro Colorado Island, Panama; expanding to Madre de Dios, Peru People: Yves Basset, David Rolnick, Maxim Larrivée Date: July 13, 2026 Dek: The forest is the library and every unnamed insect is an unread book. On a Panama island, a 17-year count meets a closing funding window, and a new set of eyes reads faster than time runs out. - **The person and the place:** For seventeen years, Yves Basset has led the arthropod count on Barro Colorado Island, Panama, cataloguing insects most of the world will never think to name. Maxim Larrivée, who leads the Montreal Insectarium, wanted to find the species no one has ever catalogued, the "dark diversity" hiding in plain sight. He brought in David Rolnick, a researcher at McGill University and the Quebec AI Institute (Mila), to build the eyes that could help him look. - **The problem the team cared about:** Basset's own count shows how much of this world is unnamed. Mongabay Latam reports that 60 percent of the insects collected on the island have no scientific name, a figure it attributes to Basset. Insects are vanishing at a pace that outstrips anything scientists have measured before, and no team of hands, however patient, can count fast enough to keep up. - **The agency moment:** Rather than continuing to rely on collection alone, the team put AMI camera-and-light-trap stations into the forest to photograph moths at scale, and Rolnick designed the open-source Antenna platform to read what the traps bring back and flag whatever the model cannot identify for a human expert to examine. The tool did not replace the naming. It widened how much ground two people could cover before the window closed. - **What changed:** In an early test of the sensors in Panama's rainforest, the team found three hundred new species within a week, Rolnick told AFP in November 2024. On Barro Colorado today, roughly a quarter of the moths photographed are still unknown to science, and the same approach is expanding into Peru's Madre de Dios, where unnamed species can make up more than half of what turns up. Meanwhile, federal funding for the island's monitoring effort has run dry, threatening to shut the program down by the end of 2026 unless new private support arrives. - **The move worth copying:** Nobody reading this is going to build a rainforest sensor network. The copyable part is smaller: they started counting before they knew what they would find. Know someone counting something nobody else thinks is worth counting? Feed the scout at goodinprogress.org. - **Takeaway line:** "Estamos quemando la biblioteca antes de leer los libros." (We are burning the library before we read the books.) — Maxim Larrivée, director, Montreal Insectarium, to Mongabay Latam Sources: - Mongabay Latam, May 2026: https://es.mongabay.com/2026/05/ultimo-refugio-polillas-inteligencia-artificial-descifrar-diversidad-oscura-panama/ - AFP wire, via Jacaranda FM, November 2024: https://www.jacarandafm.com/news/news/aiprojecthopeshelpreversemass/ - Smithsonian Tropical Research Institute, the institution that runs the Barro Colorado program, September 2025: https://stri.si.edu/story/meet-ami Sources last verified: August 19, 2026 --- ## Heather Minor didn't just fix a job she dreaded. She tells clerks across the country how to fix theirs. URL: https://goodinprogress.org/stories/heather-minor-clerkminutes-longview Location: Long View, NC, USA (pop. ~5,300) People: Heather Minor Date: July 14, 2026 Dek: For years, compiling her town's meeting minutes was one of Heather Minor's most dreaded tasks. She fixed it with an AI tool, and now she recommends the fix to clerks across the country. - **The person and the place:** Heather Minor is town clerk for Long View, North Carolina, a town of about 5,300 people. She is also the town's public relations director, two jobs held by one person. - **The problem Heather cared about:** Compiling official meeting minutes was, in her words, "one of my most daunting tasks," something she "dreaded." She would record a full council meeting, then sit through the whole recording again, stopping every 15 to 30 seconds to type out what she heard. The task ate 4 to 8 hours per meeting on top of everything else on her plate. - **The agency moment:** Rather than treat the drudgery as a permanent cost of the job, Minor brought HeyGov's ClerkMinutes into her own workflow and made it part of how she works. - **What changed:** The same task now takes her about 30 to 60 minutes instead of 4 to 8 hours. After one recent budget meeting, the tool got "every budget number correct," cutting the kind of transcription error that comes from typing numbers by ear. Minor didn't keep the fix to herself. In the company's own launch announcement, she said, "ClerkMinutes has been a true time-saver for me, and I recommend it often to other Clerks across the country." - **The move worth copying:** She took the task she dreaded most and pointed a tool at it, then told other clerks how. If something on your own calendar eats an afternoon every time it comes round, that is the one to try this on first. - **Takeaway line:** "[It got] every budget number correct. It picked up everything exactly how I want it to look." — Heather Minor, to Smart Cities Dive Sources: - Smart Cities Dive: https://www.smartcitiesdive.com/news/ai-municipal-clerks-speed-up-public-records-reporting/820216/ - Government Technology / GovTech, April 2026: https://www.govtech.com/artificial-intelligence/ai-takes-the-drudgery-out-of-compiling-meeting-minutes - WSOC-TV, Charlotte, April 2026: https://www.wsoctv.com/news/local/city-town-clerk-jobs-shift-new-ai-tool-enters-scene/L5IUSE2YPJAIXIMWMEQYMSOE2I/ - HeyGov/ClerkMinutes press release via PR Newswire, the company's own announcement, April 2026: https://www.prnewswire.com/news-releases/clerkminutes-launching-national-meeting-minutes-day-marking-the-largest-single-day-ai-deployment-in-local-government-302758167.html Sources last verified: August 5, 2026 --- ## Helen got her eye contact back. Two directors made sure she could keep it. URL: https://goodinprogress.org/stories/karen-fuller-kingston-oxfordshire-ai-notes Location: Royal Borough of Kingston upon Thames & Oxfordshire County Council, England, UK People: Karen Fuller, Sarah Spence, Helen Date: July 14, 2026 Dek: At Kingston and Oxfordshire, two directors chose to give their social workers back the time, and the attention, that paperwork was taking from the people in front of them. - **The person and the place:** On the frontline at the Royal Borough of Kingston upon Thames, England, is Helen, a senior social worker. The decision that changed her workday came from two directors: Karen Fuller, Director of Adult Social Care at Oxfordshire County Council, and Sarah Spence, Interim Director of Adult Social Care Transformation and Improvement at Kingston. - **The problem Karen cared about:** Social workers were spending more time on paperwork than on the people they were there to help. Fuller put it plainly: "We were hearing constantly that administrative work was taking longer than direct work, with people feeling torn between being fully present with residents and capturing everything accurately at the same time." For Helen, that meant writing notes during a visit instead of looking at the person she was visiting. - **The agency moment:** Fuller and Spence each led their councils into adopting Beam's AI note-taking tool, branded Magic Notes at Kingston, choosing it specifically to cut documentation time and free up staff attention. Kingston's rollout was made voluntary by design, something Spence said mattered for real adoption to take hold. - **What changed:** At Kingston, a pilot cut administrative time by more than half, and freed up to 4 hours a week for some social workers, strong enough results that the tool expanded to 100 more workers. Kingston's own case study for the Local Government Association adds that supervision write-ups fell from 40 minutes to under 10. At Oxfordshire, three quarters of the test group said the tool had done its job, with work that used to take up to eight hours often finished in more like two, and the county has since handed the tool to more than 1,800 staff in children's services, education, and adult social care. For Helen, the change wasn't measured in a percentage. It was measured in attention. - **The move worth copying:** The move was not buying software. It was deciding what the saved time was for, and then protecting it. Share this with anyone managing people whose paperwork is eating the job they came to do. - **Takeaway line:** "When I am completing an assessment, review or any other interaction with people I support, Magic Notes allows me to maintain eye contact, listen accurately and critically reflect in the moment and it means that I don't have to worry about taking notes." — Helen, Senior Social Worker, Kingston Sources: - Diginomica, June 2026: https://diginomica.com/how-two-uk-councils-are-using-ai-note-taking-cut-social-workers-admin-burden - Kingston Nub News, November 2024: https://kingston.nub.news/news/local-news/kingston-council-gives-its-social-workers-access-to-ai-tool-245678 - Beam, the company that makes the tool, case study, no date given: https://beam.org/en-us/case-studies/kingston-and-beam-case-study - Local Government Association case study, written by Kingston Council, no date given: https://www.local.gov.uk/case-studies/kingston-council-using-ai-adult-social-care-administration Sources last verified: August 13, 2026 --- ## He wrote himself a song in a voice his illness had taken URL: https://goodinprogress.org/stories/patrick-darling-elevenlabs-singing-voice Location: Bristol, England, UK (performance at the ElevenLabs Summit, London, Feb 11 2026) People: Patrick Darling, Richard Cave, Nick Cocking, Hari Ma Date: July 14, 2026 Dek: Motor neurone disease took Patrick Darling's ability to sing. Working with a speech and language therapist and AI voice technology, he rebuilt his singing voice from old recordings, wrote an original song in his own words, and sat on stage in London as it played, his old bandmates playing live beside him. - **The person and the place:** Patrick Darling, 32, of Bristol, England, was the singer and guitarist for The Ceili House Band, an Irish folk group, until motor neurone disease took his voice. - **The problem Patrick cared about:** Diagnosed at 29, Darling lost his legs first, then the ability to stand or play his instruments, and finally the ability to sing or speak clearly. "It felt like I was losing an essential part of my identity," he said in ElevenLabs' own account of the project. - **The agency moment:** Darling worked with Richard Cave, a speech and language therapist and University College London researcher who is also a consultant for ElevenLabs, using recordings from before his diagnosis to rebuild a version of his singing voice. He wrote an entirely new song around it in his own words, "Ghost of a Man I Never Met," for the great-grandfather he never got the chance to meet, and he and Cave spent around six weeks fine-tuning it. - **What changed:** On February 11, 2026, the song played while Darling sat on stage in London, with bandmates Nick Cocking and Hari Ma playing their parts live alongside it, in front of an audience ElevenLabs put at nearly 1,000 people. "It sounded exactly like I had before, and you literally wouldn't be able to tell the difference," he said at the event, speaking through the voice clone. His family watched from the audience. Until that night, they hadn't heard him sing since the disease took hold. - **The move worth copying:** There is nothing to copy here, and that is alright. He wanted to sing with his band one more time, and he did. Share this with someone who thinks their window has already closed. - **Takeaway line:** "The first time I heard the voice, I thought it was amazing... It sounded exactly like I had before, and you literally wouldn't be able to tell the difference." — Patrick Darling, on stage at the ElevenLabs Summit, reported by MIT Technology Review Sources: - MIT Technology Review, February 2026: https://www.technologyreview.com/2026/02/13/1132913/als-stole-this-musicians-voice-ai-sing/ - dpa/PA Media wire, February 2026, via The Star, and the same wire copy also ran at The Independent and Neuro Rehab Times: https://www.thestar.com.my/tech/tech-news/2026/02/25/ai-helps-musician-sing-again-after-he-lost-voice-to-disease - Harbingers' Magazine, March 2026: https://hrbmagazine.org/articles/inside-the-elevenlabs-summit-where-ai-found-its-voice-by-charlotte-wejchert/ - ElevenLabs company blog, the company behind the voice tool, February 2026: https://elevenlabs.io/blog/when-humanity-meets-technology Sources last verified: August 19, 2026 --- ## Jorge Bastidas didn't wait for the government. He and his team built the missing-persons list in three hours. URL: https://goodinprogress.org/stories/jorge-bastidas-desaparecidos-venezuela Location: Buenos Aires, Argentina (built for Venezuela, after the June 2026 earthquakes) People: Jorge Bastidas Date: July 16, 2026 Dek: After twin earthquakes hit Venezuela and official reports stalled, a 31-year-old programmer in Buenos Aires built a citizen-run missing-persons tracker. He says he and a team of six had it running in three hours, and that it took in more than 30,000 reports in its first two days. - **The person and the place:** Jorge Bastidas is a 31-year-old Venezuelan programmer living in Buenos Aires, Argentina. - **The problem Jorge cared about:** Twin earthquakes struck northern Venezuela on June 24, 2026. Families searching for missing loved ones needed answers, and the government's official reporting was slow to arrive. - **The agency moment:** Bastidas decided not to wait for an official response. "Who knew when the government was going to respond? We decided to take action," he said. Working with Claude Opus 4.8 and donated facial-recognition software, he says he and a team of six built Desaparecidos Terremoto Venezuela, a citizen-run missing-persons tracker, in three hours. Without AI, "it would have taken me about 24 hours without rest to build something that took me three hours," he said. - **What changed:** In its first two days online, Bastidas told Rest of World, the site took in more than 30,000 reports of missing people. - **The move worth copying:** He didn't wait for a press release or a permission slip. He opened his laptop the night his country needed it and built the thing himself. If something where you live needs a list nobody is making, three hours is less time than you think. - **Takeaway line:** "Who knew when the government was going to respond? We decided to take action." — Jorge Bastidas, Rest of World Sources: - Rest of World: https://restofworld.org/2026/venezuela-ai-citizen-disaster-response/ - France 24: https://www.france24.com/es/am%C3%A9rica-latina/20260627-listas-en-la-pared-y-webs-caseras-los-venezolanos-buscan-a-sus-desaparecidos-tras-los-terremotos - El Nacional: https://www.elnacional.com/2026/06/red-ciudadana-digital-emerge-para-localizar-desaparecidos-tras-los-terremotos/ - Desaparecidos Terremoto Venezuela, the project's own self-published press page: https://desaparecidosterremotovenezuela.com/medios/ Sources last verified: September 5, 2026 --- ## Michael Running Wolf left Amazon so Indigenous tribes could own their own AI. URL: https://goodinprogress.org/stories/michael-running-wolf-flair Location: Montreal, Quebec, Canada (McGill University); FLAIR operates across North America People: Michael Running Wolf Date: July 16, 2026 Dek: Roughly 200 Indigenous languages in North America are endangered, and the AI built to help save them was being built by outsiders. Michael Running Wolf left Amazon's Alexa team to make sure tribes controlled it instead. - **The person and the place:** Michael Running Wolf describes himself as Northern Cheyenne, Lakota, and Blackfeet. He grew up in Montana and went on to a computer science PhD at McGill University in Montreal. - **The problem Michael cared about:** Roughly 200 Indigenous languages across North America are endangered. The speech and language AI being built to help preserve them was largely being built, owned, and controlled by companies outside those communities. - **The agency moment:** Running Wolf left his job on Amazon's Alexa team, where he'd founded the company's first Indigenous employee affinity group, to co-found and lead FLAIR (First Languages AI Reality). His goal wasn't to build the one tool that saves every language himself. "I want there to be an ecosystem of research. I don't want us to be the only ones doing this. I don't want to have control over their AI destiny. I want tribes to take over their own destiny," he said. - **What changed:** The Lakota AI Code Camp, which Running Wolf co-founded, took in 20 rural South Dakota kids with no prior software experience. Every camper who finished went on to a four-year institution, studying computer science, maths, or something close to it. Back in 2023 he described a further-out hope for the technology itself: "I want to put on a headset, and be able to speak Lakota and experience a Lakota buffalo hunt or engage with conversations with digital avatars or chat bots in VR as an educational experience." By 2025 he had pushed that horizon out. "I think that was very pie-in-the sky, obviously. That's not going to happen any time soon; there are a lot of technical barriers. Our near-time objective is very modest," he said. - **The move worth copying:** He didn't just want to save the language. He wanted to make sure the people who speak it still own the tools that do it. If your community is being offered help that comes with someone else holding the keys, this is the version worth pointing at. - **Takeaway line:** "I want there to be an ecosystem of research." — Michael Running Wolf, Prism Reports Sources: - Prism Reports: https://prismreports.org/2026/02/26/indigenous-languages-preservation-ai/ - Modern Languages Open, an interview Running Wolf co-authored: https://modernlanguagesopen.org/articles/10.3828/mlo.v0i0.578 - Voices of VR podcast, a 2023 interview with Running Wolf: https://voicesofvr.com/1258-using-xr-ai-to-reclaim-and-preserve-indigenous-languages-with-michael-running-wolf/ Sources last verified: August 14, 2026 --- ## Jason Eshraghian and Rui-Jie Zhu built an AI that runs on the power of a lightbulb. URL: https://goodinprogress.org/stories/jason-eshraghian-matmul-free-llm Location: Santa Cruz, CA, USA People: Jason Eshraghian, Rui-Jie Zhu Date: July 16, 2026 Dek: Large language models are notorious for their energy appetite. A UC Santa Cruz professor and his graduate student found a way to strip the most expensive step out of the math, and in 2024 ran a model with a billion parameters on 13 watts, about the draw of a lightbulb. - **The person and the place:** Jason Eshraghian, assistant professor of electrical and computer engineering at UC Santa Cruz's Baskin School of Engineering, and Rui-Jie Zhu, his graduate student and the paper's first author. - **The problem Jason and Rui-Jie cared about:** Running large language models costs real money and real energy, and Eshraghian wanted to know whether that cost was actually necessary. "The idea started by acknowledging the fact that language models like ChatGPT are incredibly expensive in terms of the amount of resources that you need to run them," he said. - **The agency moment:** In 2024, instead of accepting that bigger, pricier hardware was the only path forward, Eshraghian and Zhu set out to eliminate matrix multiplication, the single most computationally expensive step in running a neural network, and then built custom hardware to prove it could be done. "We got the same performance at way less cost — all we had to do was fundamentally change how neural networks work. Then we took it a step further and built custom hardware," Eshraghian said. They built the low-power prototype system in three weeks. - **What changed:** Their custom board ran a model with a billion parameters while drawing 13 watts. The same job on GPUs takes roughly 700 watts, which makes their board more than fifty times as efficient. On standard GPUs, the same approach cut memory use about tenfold and ran about 25% faster, matching the performance of language models its own size. That result went out in 2024 as a preprint (arXiv 2406.02528) rather than a peer-reviewed paper, and the comparison is to models at the same scale, not to systems like GPT-4, which are estimated to carry more than a trillion parameters. - **The move worth copying:** They didn't wait for cheaper chips to make AI less expensive to run. They took the priciest part of the math out of the machine. The copyable part is the question, not the chip design: which expensive step is everyone treating as fixed? Know someone asking that about their own field? Feed the scout at goodinprogress.org. - **Takeaway line:** "We are just a small academic lab that started less than two years ago and we are capable of competing with the giants." — Jason Eshraghian, Santa Cruz Sentinel via GovTech Insider Sources: - UC Santa Cruz News, the researchers' own university newsroom, June 2024: https://news.ucsc.edu/2024/06/matmul-free-llm/ - Santa Cruz Sentinel, distributed by Tribune Content Agency and read via GovTech Insider, July 2024: https://insider.govtech.com/california/news/uc-santa-cruz-researchers-want-to-scale-back-ais-carbon-footprint - Digital Trends, June 2024: https://www.digitaltrends.com/computing/researchers-build-ai-use-100w-light-bulb-power/ Sources last verified: August 10, 2026 --- ## The Cambridge technologist who built a $50 device to save a language with no written script URL: https://goodinprogress.org/stories/saving-voices-soliga Location: Southern India (Soliga Indigenous community); project led from Cambridge University, UK and IIIT Dharwad, India People: Arjuna Sathiaseelan, Sarbani Banerjee Belur Date: July 16, 2026 Dek: Soliga has no written script, and as young people left for city jobs the elders worried it would go quiet. A Cambridge University technologist and an IIIT Dharwad professor built the community a working voice model from five hours of recordings, running offline on hardware that costs less than $50. - **The person and the place:** Arjuna Sathiaseelan, founder of the Saving Voices Project and CTO of the Frugal AI Hub at Cambridge University, working with Sarbani Banerjee Belur, co-founder and assistant professor at IIIT Dharwad, alongside the Soliga Indigenous community in southern India. - **The problem Arjuna and Sarbani cared about:** Soliga has no written script. Young people were leaving for city jobs, and the elders could see what that meant for a language that lives only in speech. Commercial speech technology was never built with a language like this in mind. - **The agency moment:** Rather than wait for a commercial speech-AI vendor to take an interest in an unwritten Indigenous language, Sathiaseelan and Belur built a custom offline text-to-speech system from scratch, using just five hours of recorded Soliga voice. The recordings never left the community's own devices, "something that simply isn't possible with closed cloud systems," Sathiaseelan said. - **What changed:** The model lives on a Raspberry Pi that costs less than $50 and keeps working offline for long stretches. The word error rate in these models still runs a little high, but "what we gained is complete data sovereignty, offline deployment on sub-$50 hardware, and a governance structure that elders and community leaders actually trust," Sathiaseelan said. - **The move worth copying:** They didn't wait for Big Tech to make room for a language with no written script. They built the room themselves, for less than the cost of a dinner out. A Raspberry Pi and under fifty dollars. The bar for building something for a language, or a place, that big tech has skipped is lower than it looks. - **Takeaway line:** "With just five hours of voice data, we were able to build a voice model for the Soliga by prioritizing community ownership, and with frugal, deployable technology." — Arjuna Sathiaseelan, Rest of World Sources: - Rest of World: https://restofworld.org/2026/frugal-ai-big-tech/ - Silicon Canals, a commentary column drawn from the Rest of World report: https://siliconcanals.com/sc-d-the-global-south-is-building-its-own-ai-on-50-hardware-and-its-working/ Sources last verified: August 20, 2026 --- ## The city attorney who used AI to find the reports nobody was reading URL: https://goodinprogress.org/stories/sf-city-attorney-ai-code-declutter Location: San Francisco, CA, USA People: David Chiu, Daniel Ho, Andrea Bruss Date: July 16, 2026 Dek: San Francisco's 16-million-word municipal code had quietly grown hundreds of mandatory reports that almost no one read. City Attorney David Chiu and Stanford's RegLab built a tool to find them all, made his own staff check every one by hand, and the Board of Supervisors just voted on what to do about it. - **The person and the place:** David Chiu, City Attorney of San Francisco, working with Daniel Ho, Faculty Director of Stanford's Regulation, Evaluation, and Governance Lab. - **The problem David cared about:** the city's code had swelled to nearly 16 million words and more than doubled its mandatory reporting requirements since 2000, to the point that, in Chiu's own words, "that makes systemic reform daunting." - **The agency moment:** Instead of accepting the code as unreadable, Chiu's office partnered with RegLab, which pointed a legal-research AI tool at the entire code to surface every reporting requirement buried in it. "The power of this technology is the ability to do something like this at scale, in a way that just would be infeasible for humans to do manually," Ho said. Then Chiu's own lawyers checked the machine's work: Director of Government Legal Reform Andrea Bruss said the tool "did not replace the need for us to individually evaluate the legal issues associated with every single one of these outputs." - **What changed:** The tool surfaced 528 reporting requirements, of which 488 could legally be changed. Chiu's office introduced a 351-page ordinance in June 2025 proposing to delete, modify, or consolidate 174 of them, 36% of everything on the table. The bill grew to 362 pages after a year of amendments, and on July 14, 2026 the Board of Supervisors passed it 7-4, with Supervisors Myrna Melgar, Connie Chan, Jackie Fielder, and Shamann Walton voting no. In the final two weeks, amendments restored several requirements the AI review had flagged for removal, including domestic-violence data sharing, sexual-harassment reporting, and small-business case-manager reporting, after named officials raised concerns; a surveillance-technology audit requirement was changed from annual to every five years over Supervisor Fielder's objection. - **The move worth copying:** The code got too big to read by hand, so the city attorney built something that could read it, made sure a human checked every page, and then let the elected body argue over what stayed. That division of labour is the part worth copying: let the machine read, keep a human on every page, and leave the arguing to the people elected to do it. - **Takeaway line:** Chiu said he hopes San Francisco's code cleanup "serves as a model for us and other governments." — David Chiu, City Attorney of San Francisco, to the San Francisco Examiner Sources: - Mission Local: https://missionlocal.org/2026/07/sf-city-code-streamlining-mandelman/ - San Francisco Examiner: https://www.sfexaminer.com/news/the-city/sf-city-attorney-turns-to-ai-to-cut-down-on-policy-sludge/article_56f86a1c-84df-4610-b50d-e2a7accb7667.html - Mission Local, an earlier report from the same outlet: https://missionlocal.org/2026/06/s-f-supervisors-balk-at-362-page-city-code-rewrite/ - City Attorney of San Francisco press release, a primary document republished verbatim by Stanford Law School: https://www.sf.gov/news-city-attorney-introduces-legislation-to-modernize-municipal-code-with-technology - Reboot Democracy, an advocacy series that champions the project, background: https://rebootdemocracy.ai/blog/sf-ai-regulatory-decluttering - Niskanen Center, a promotional listing for an event featuring the project's own participants, background: https://www.niskanencenter.org/events/virtual-event-modernizing-municipal-code-with-technology-how-the-city-of-san-francisco-used-ai-to-cut-red-tape/ Sources last verified: August 24, 2026 --- ## Daniela Camberos built the AI that checks a job offer before it can become a trafficking trap. URL: https://goodinprogress.org/stories/daniela-camberos-ramona-ia Location: Mexico People: Daniela Camberos Date: July 16, 2026 Dek: Criminal networks use fake job listings as the first hook to recruit trafficking victims. Daniela Camberos, a Mexican HR specialist, built a free AI bot that checks the offer in real time, before anyone walks into the trap. - **The person and the place:** Daniela Camberos is a Mexican human-resources specialist and the founder of Ramona IA. - **The problem Daniela cared about:** Criminal networks use fraudulent job offers as their first hook to recruit trafficking victims, especially women. Camberos put the gap in her own words: "RAMONA was born from a dual problem I identified in human resources: the total absence of mechanisms to validate legitimate job offers and the lack of access to career guidance for young people, especially in public universities and vulnerable communities." - **The agency moment:** Rather than wait for an employer, a platform, or a government program to close that gap, Camberos built Ramona IA, a free chat bot that runs 24/7, works on low-end devices and limited connectivity, and pairs fraud detection with free job coaching, CV help, interview prep, and labor-rights education. She built the first version for Telegram, and she now leads a team of engineers and employability specialists who run it on WhatsApp. - **What changed:** The system is reported to catch fraudulent job offers with close to 98% precision. Cimacnoticias reports the tool has protected more than 800 people from potential fraud, and that the startup has contributed to the empowerment of more than 2,000 women in Mexico. Camberos has drawn recognition of her own: the Geneva Centre for Human Rights Advancement gave her its Individual Award for Ethical AI and Human Rights Protection in December 2025. - **The move worth copying:** She didn't wait for someone else to build the checkpoint. She built it herself, made it free, and gave it to the people who needed it most. Know someone about to take a job offer from a stranger? Send them this. - **Takeaway line:** "Criminal networks use job offers as the first hook to recruit victims. Ramona validates in real time whether a job offer is legitimate or fraudulent..." — Daniela Camberos, to Contxto Sources: - Cimacnoticias: https://cimacnoticias.com.mx/2025/12/15/mexicana-crea-tecnologia-para-combatir-fraude-laboral-digital-ramona-ia/ - Contxto, an interview with Camberos: https://contxto.com/en/entrevistas/ramona-ia-the-ai-tool-created-by-mexican-entrepreneur-to-protect-young-people-from-fraudulent-job-offers/ - El Ecosistema Startup, which lists Contxto as its 'fuente original' in its own Fuentes section; not independent corroboration: https://ecosistemastartup.com/ramona-ia-la-startup-mexicana-que-combate-el-fraude-laboral-con-ia/ - Geneva Centre for Human Rights Advancement and Global Dialogue, the awarding body: https://gchragd.org/geneva-centre-holds-fifth-edition-of-award-focused-on-human-rights-in-the-digital-age/ - UNESCO AI Ethics Observatory directory, Ramona's own self-submitted profile: https://www.unesco.org/ethics-ai/en/civil-society-organizations/ramona-ai - Ramona's own company site: https://ramonaaliadalaboral.com/ Sources last verified: October 4, 2026 --- ## His grandfather told him stories about the rain prophets. Raul Souza taught a machine to listen. URL: https://goodinprogress.org/stories/raul-souza-profetas-da-chuva Location: Jaguaribe Valley, Ceará, Brazil (sertão) People: Raul Victor Magalhães Souza Date: July 19, 2026 Dek: A 16-year-old in Ceará wrote down what six of Brazil's traditional rain prophets watch for, from the flowering of the mandacaru to the brightness of the Sete-Estrelo, crossed it with weather records going back to 1981, and tested the result against a real rainy season. - **The person and the place:** Raul Victor Magalhães Souza is 16 and studies at a state public high school in Fortaleza, Ceará. The rain prophets he went looking for are in the Jaguaribe Valley, in Brazil's semi-arid sertão. - **The problem Raul cared about:** In the valley the rainy quarter decides the year, and the rain is not dependable. Water is scarce, the season is uncertain, and drought lands directly on the community that lives there. The people who have read that season locally for generations are the rain prophets, who watch for a lunar halo, the flowering of the mandacaru, a black butterfly, the brightness of the Sete-Estrelo, and call the coming rains from what they see. Raul's grandfather, the farmer Luiz Maia, was his way in. It was his grandfather's stories about the prophets that gave him the idea for the research. - **The agency moment:** Raul treated the prophets' knowledge as data nobody had bothered to write down, so he wrote it down. Six prophets across five municipalities told him what they watch for, and he turned their answers into ten measurable parameters covering animal behavior, plant signs, and atmospheric conditions. Against those he set the Funceme and Inmet records from 1981 to 2024, more than 40,000 of them. His professor, Helyson Lucas Bezerra Braz, mentored the project, a UFC laboratory partnered on it, and per Diário do Nordeste the two of them built the result into a machine-learning model behind a Streamlit interface. - **What changed:** They tested the system, Inteligência Artificial dos Profetas da Chuva, against the 2025 rainy season. Between January and March it estimated 422.1 mm of rain; 431.5 mm actually fell, a mean monthly error of 5.7%. For comparison, a commercial reference platform put the same season at 318 mm, off by 30.8%. ABES, which organizes the Brazilian stage of the Stockholm Junior Water Prize, also reports 94.5% accuracy in the model's classification performance, from that one season of validation. The project won the high-school category of Brazil's 31st Prêmio Jovem Cientista, then the Brazil stage of the Stockholm Junior Water Prize. In August 2026 Raul went to the international final in Stockholm during World Water Week as Brazil's representative. "Tenho que agradecer ao meu avô, minha principal referência e base para a construção do projeto," he said. I have to thank my grandfather, my main reference and the foundation for building this project. - **The move worth copying:** He didn't choose between his grandfather's knowledge and the data. He wrote down what the elders already knew, carefully enough that a machine could learn it. If you know someone whose knowledge lives only in their head, writing it down is the whole first step. - **Takeaway line:** "Tenho que agradecer ao meu avô, minha principal referência e base para a construção do projeto." (I have to thank my grandfather, my main reference and the foundation for building this project.) — Raul Victor Magalhães Souza, ABES Sources: - Diário do Nordeste: https://diariodonordeste.verdesmares.com.br/ceara/estudante-cearense-vence-premio-nacional-com-trabalho-sobre-profetas-da-chuva-e-ia-1.3713251 - GZH / Zero Hora, original reporting by Rafael Parente, read here via NIC.br's republication: https://nic.br/noticia/na-midia/estudante-vence-premio-jovem-cientista-ao-integrar-ia-a-saberes-ancestrais/ - ABES, which organizes the Brazilian stage of the prize, announcing its own winner: https://www.abes-dn.org.br/estudante-do-ceara-e-o-grande-vencedor-da-etapa-brasileira-do-stockholm-junior-water-prize-2026/ Sources last verified: August 24, 2026 --- ## The security researcher who built his own sky-searching tools, then found a world that glows URL: https://goodinprogress.org/stories/dan-caselden-brown-dwarfs Location: Massachusetts, USA People: Dan Caselden, Paul Westin, Adam Schneider Date: July 20, 2026 Dek: Dan Caselden looks for cold, starless worlds in his spare time. The one he found turned out to be glowing, and when the team reported it in January 2024 the cause was still an open question. - **The person and the place:** By day, Dan Caselden is a computer security researcher. In his own hours he is a citizen scientist with Backyard Worlds: Planet 9, a NASA-funded project that invites anyone at all to comb through spacecraft data for brown dwarfs, the objects too large to be planets and too small to be stars. - **The problem Dan cared about:** Brown dwarfs are cold, faint, and very good at hiding. Volunteers work through image after image by eye, because a thing that moved is still something the human eye catches best. "They're really hard to find," Caselden says. "It's the needle in the haystack problem, which is so much fun." - **The agency moment:** He did not just click through the images he was handed. With his colleague Paul Westin he built WiseView, which animates the survey images so that anything moving gives itself away, and he went on to write machine learning code for the search. Astronomer Adam Schneider used WiseView to confirm that an object he had noticed years earlier was moving fast. - **What changed, with evidence:** Through the project, Caselden found W1935, bigger than Jupiter, too small to be a star, drifting with no sun of its own. When Jackie Faherty's team pointed the James Webb Space Telescope at it, the methane in its atmosphere was not absorbing light the way everyone expected. It was glowing, the signature of an aurora, on an object with no star to spark one. Faherty: "We expected to see methane because methane is all over these brown dwarfs. But instead of absorbing light, we saw just the opposite: The methane was glowing." When the team reported it in January 2024, the cause was still open. Their own candidates were an internal process, interstellar plasma, or an unseen active moon, and none was confirmed. - **Where it led:** Caselden is now a co-investigator on a Webb observing campaign studying these cold worlds. Curiosity was the whole qualification; he brought the tools. - **The move worth copying:** Curiosity was the entire qualification; he brought the tools himself. Backyard Worlds is still open, and the next cold world in the data is waiting for whoever is patient enough to notice it. - **Takeaway line:** "It's chock full of really hard problems. And if you make any improvement, it's felt by the science team. You can see that what you do has an actual impact." — Dan Caselden, Discover Magazine Sources: - Discover Magazine, a Science Near Me partner post promoting the project, 2022: https://discovermagazine.com/the-sciences/citizen-scientists-are-using-the-jwst-to-study-mysterious-brown-dwarfs - Earth.com, January 2024: https://www.earth.com/news/cosmic-enigma-aurorae-discovered-on-isolated-brown-dwarf/ - Universe Today, January 2024: https://www.universetoday.com/165156/a-solo-brown-dwarf-found-with-auroras/ - NASA/JPL, the project's funder, July 2020: https://www.jpl.nasa.gov/news/two-bizarre-brown-dwarfs-found-with-citizen-scientists-help/ Sources last verified: October 4, 2026 --- ## He drove thirty minutes at five in the morning to read a thermometer for five seconds. Then Hiroki Tomiyasu built the thing that reads it for him. URL: https://goodinprogress.org/stories/hiroki-tomiyasu-farm-ai Location: Biratori and Mukawa, Hokkaido, Japan People: Hiroki Tomiyasu Date: July 21, 2026 Dek: A liberal-arts graduate who had never written a line of code taught himself to build his own farm software, and now a 100-hectare farm in Hokkaido leans on tools that started with about 7,000 yen of consumer sensors and a winter of evenings. - **The person and the place:** Hiroki Tomiyasu came north to Hokkaido as a member of a regional revitalization cooperation squad, and started farming in 2020. He is now an executive at WFP Dacho Farm, about 100 hectares across the towns of Biratori and Mukawa, growing broccoli, soybeans, squash, and green onions. - **The problem Hiroki cared about:** The farm kept growing and the crew did not. 「1棟だと知れているんですが、20棟～30棟になってくると移動も含めて大変」, he told HBC. One greenhouse is a known quantity, but once it is twenty or thirty, it gets hard, the travelling included. The smallest chore was the worst of them. He described his old mornings to Metagri Labo: 「朝5時。エンジンをかけて車が暖まるまで5分。ハウスまで10分。温度計を覗く。『3.2℃か…大丈夫だな』。そのまま帰る。往復30分、確認5秒。」 Five in the morning. Five minutes to warm the car, ten to the greenhouse, a glance at the thermometer, 3.2 degrees, we're fine, and drive home. Thirty minutes round trip for five seconds of looking. - **The agency moment:** He is a liberal-arts graduate from Kanagawa with no programming education at all. In the winter of 2025 he started teaching himself Python, then started asking AI instead of a textbook. From day nine he posted one small app a day in public, which left him nowhere to quietly quit. - **What changed, with evidence:** He built a greenhouse temperature monitor out of consumer sensors for about 7,000 yen up front, running at zero yen a month, where commercial farm IoT services typically run several thousand to several tens of thousands of yen a month, by Metagri Labo's accounting. He built a remote window opener from a small controller and a motor over about two months, for 60,000 to 70,000 yen in parts. 「外注すれば100万円程度はかかるかもしれません」, he told Business Insider Japan: outsourced, it might have run about a million yen. He built a GPS work logger in roughly a day. He built a bot that carries Japanese into English, Indonesian, and Khmer for the people working the fields alongside him. HBC reports that a simple program takes him about fifteen minutes to get working, at roughly a tenth of what outsourcing costs. - **Where it led:** He keeps building for a reason that has little to do with money. 「自作すると、温度、操作時刻、開閉履歴、手動か自動か、失敗ログまで自分で持てます。」 Build it yourself, he writes, and you hold all of it: temperature, operating times, open-close history, manual or automatic, even the failure logs. The tools fit his fields because he is the one who knows the fields. Anyone wondering what their own version might be can start where he started, with the chore they keep driving out to do by hand. - **The move worth copying:** Seven thousand yen and consumer sensors. Build it yourself, he writes, and you keep the temperatures, the timings, and the failure logs. Anyone with a problem they understand better than any vendor could start where he started. - **Takeaway line:** "チャットGPTで会話するような感覚でシステムを作っていける。トラクターの中で思いついて、トラクターの中から指示を出して、スマホがリモコンになってパソコンが勝手に作業しているみたいな" (You can build a system with the feeling of having a conversation with ChatGPT. I think of it inside the tractor, give the instruction from inside the tractor, and it is as if my phone became the remote and the computer is off doing the work on its own.) — Hiroki Tomiyasu, Sitakke (HBC Hokkaido Broadcasting) Sources: - note, Tomiyasu's own author page: https://note.com/tomiyasu16 - note, Tomiyasu's own post, 'Build the system yourself and the field data becomes yours', May 2026: https://note.com/tomiyasu16/n/n257919cad20d - Business Insider Japan, June 2026: https://www.businessinsider.jp/article/2606-openai-codex-farm-automation/ - Sitakke, the web magazine of HBC Hokkaido Broadcasting, June 2026: https://sitakke.jp/post/20493/ - Hokkaido Shimbun, paywalled past the opening paragraphs, June 2026: https://www.hokkaido-np.co.jp/article/1323480/ - Nogyo AI Tsushin / Metagri Labo, an interview with Tomiyasu, Vol. 1: the 100-day app challenge: https://metagri-labo.com/ai-guide/tomiyasu-ai-interview-vol1/ - Nogyo AI Tsushin / Metagri Labo, an interview with Tomiyasu, Vol. 2: the zero-yen-a-month greenhouse monitor: https://metagri-labo.com/ai-guide/tomiyasu-ai-interview-vol2/ - Nogyo AI Tsushin / Metagri Labo, an interview with Tomiyasu, Vol. 3: running the farm's data through LINE: https://metagri-labo.com/ai-guide/tomiyasu-ai-interview-vol3/ Sources last verified: August 7, 2026 --- ## The naturalist who gave almost every humpback in the North Pacific a fingerprint URL: https://goodinprogress.org/stories/ted-cheeseman-happywhale Location: Global platform; contributors in more than 100 countries People: Ted Cheeseman, Ken Southerland Date: July 23, 2026 Dek: Ted Cheeseman built a website where anyone with a whale-tail photo can help identify individual whales, and it grew into the biggest photographic record anyone has built for a whale species. - **The person and the place:** Ted Cheeseman is a lifelong naturalist, and he is pursuing a PhD on humpback whales at Southern Cross University in Australia. - **The problem Ted cared about:** By the ITU's account, identifying one humpback from a photo of its tail could take a scientist more than an hour, and matching by hand was too slow for the ocean-scale collaboration the science needed. - **The agency moment:** In 2015 he co-founded Happywhale with Ken Southerland. Anyone can upload a photo of a whale's tail, kayakers and tourists as readily as researchers, an AI proposes a match from the unique markings, and a scientist confirms every one. - **What changed, with evidence:** It grew into the biggest photographic record anyone has built for a whale species. When the ITU gave Happywhale an AI for Good award in July 2026, it counted more than 1.5 million photos, more than 120,000 whales and dolphins identified, and 53,000 people in more than 100 countries sending pictures in. Its first big finding, in Royal Society Open Science with more than 70 co-authors, showed the North Pacific humpback population drop by roughly 7,000 whales between 2012 and 2021, close to a fifth of it, a marine-heat-wave signal no single researcher could have seen without everyone's photos. - **Where it led:** The data now helps scientists and agencies like NOAA track how a warming ocean is reshaping whale populations, built from photos taken by people who were simply out on the water. - **The move worth copying:** He built it so that a tourist with a phone counts as a contributor rather than an audience. If you have whale-tail photos sitting in an old camera roll, they are still worth uploading. - **Takeaway line:** "It's this five-meter banner, saying: This is my identity!" — Ted Cheeseman, Fast Company Sources: - Fast Company, live page blocked to readers, March 2024: https://www.fastcompany.com/91048032/inside-a-massive-ai-project-that-uncovered-7000-whale-deaths-over-9-years - NOAA Fisheries, an institutional partner whose scientists co-authored the paper, February 2024: https://www.fisheries.noaa.gov/feature-story/ai-sorts-public-photos-show-recovering-pacific-humpback-whales-hit-climate-ceiling - ITU AI for Good, the body that gave Happywhale its 2026 award, July 2026: https://aiforgood.itu.int/three-awards-one-mission-meet-the-2026-ai-for-good-impact-awards-winners/ - Courthouse News, May 2024: https://www.courthousenews.com/humpback-whale-population-data-offers-grim-prognosis-for-ocean-health/ Sources last verified: September 6, 2026 --- ## The chemistry professor who set thermodynamics to music URL: https://goodinprogress.org/stories/oliver-steinbock-chemistry-songs Location: Tallahassee, Florida, US People: Oliver Steinbock Date: July 23, 2026 Dek: Oliver Steinbock used AI to turn dense physical-chemistry equations into a playlist of songs, so the ideas would lodge in memory the way a jingle does. - **The person and the place:** Oliver Steinbock is a chemistry professor at Florida State University in Tallahassee, and he noticed how a tune sticks in your head in a way a textbook page rarely does. - **The problem Oliver cared about:** Thermodynamics is a tricky topic, and the trick of setting facts to music, common for language learners, is almost never used in the sciences. - **The agency moment:** He asked a large language model to write song lyrics for a thermodynamics topic, telling it the tempo and the style he wanted. Then he checked the lyrics for factual errors and kept adjusting his prompts until the science held up, and ran the results through the AI music tool Suno to turn each one into a song, in genres from country to K-pop, a whole playlist of them. - **What changed, with evidence:** The songs turn equations you can barely read into lines you can sing, like "T is how U answers S / When volume holds its breath" for T = (∂U/∂S)V. The workflow was written up in Nature Reviews Chemistry, and Mark Elliott, director of undergraduate studies at Cardiff University, called it "an absolute work of genius." - **Where it led:** Steinbock can picture a class assignment where students write their own songs about a hard topic, then fact-check each other's. Mark Elliott is already ahead of him. "It's so inventive and has inspired me to try something that I didn't know could be done," he said, and started wondering whether the first-row transition metals could have a song each. - **The move worth copying:** He turned equations his students could barely read into lines they could sing. If you teach something that keeps bouncing off people, the fix might not be explaining it again. - **Takeaway line:** "There's something between music and memory. Jingles and tunes are often stuck in your head for a very long time." — Oliver Steinbock, Chemical & Engineering News Sources: - Chemical & Engineering News: https://cen.acs.org/physical-chemistry/ai-thermodynamics-music-education-physical-chemistry/104/web/2026/07 - FSU News, Florida State University's own newsroom: https://news.fsu.edu/news/science-technology/2026/07/16/fsu-chemistry-professor-develops-artificial-intelligence-songs-to-help-students-learn-complex-concepts/ - Nature Reviews Chemistry, Steinbock's own paper: https://doi.org/10.1038/s41570-026-00849-0 Sources last verified: August 14, 2026 --- ## The California Teacher of the Year who turned ChatGPT into a writing coach URL: https://goodinprogress.org/stories/casey-cuny-chatgpt-writing-coach Location: Santa Clarita, California, US People: Casey Cuny Date: July 24, 2026 Dek: His students were already using ChatGPT for everything, so Casey Cuny rebuilt his English class around a routine that makes the chatbot quiz them instead of write for them. - **The person and the place:** Casey Cuny teaches 10th-grade Honors English and a senior mythology elective at Valencia High School in Santa Clarita, California. In 2024, the state named him California Teacher of the Year. - **The problem Casey cared about:** His students were already using ChatGPT, and mostly not to learn. "Some teachers are a little head-in-the-sand about it, not even realizing how much kids are using it," he told K-12 Dive. Later in the same interview: "The cheating is off the charts. I do have major worries about that." He wanted the thinking back in the room. - **The agency moment:** Instead of banning the tool, he redesigned the work around it. Students prompt ChatGPT to quiz them on a text, twenty questions, one at a time. They take their notes by hand, and once they have worked a topic through with the chatbot they shut their Chromebooks and talk it out as a class. One analog lock stays in place: "Then, I have them hand-write the thesis, so I know they wrote it." - **What changed, with evidence:** The moment that sticks came from a senior who kept pressing to finish an in-class writing assignment at home. Told no, he said, "Low-key, Mr. Cuny, this is the first thing I've written in three years." When Cuny showed his surprise, the student explained: "Yeah, I just put it into ChatGPT." - **The move worth copying:** He didn't wait for a policy. He took the tool his students already had and gave it a different job, questioner instead of ghostwriter, with one handwritten line to keep the work honest. Any teacher could run the quiz-me routine on Monday. Know one wrestling with the same wave? Share this story, or feed the scout at goodinprogress.org. - **Takeaway line:** "It bolsters critical thinking, engagement and personalization in ways we could never do before." — Casey Cuny, K-12 Dive Sources: - K-12 Dive: https://www.k12dive.com/news/first-thing-ive-written-in-3-years-students-ai-habits-prompt-teacher-t/813109/ - Associated Press: https://apnews.com/article/ai-cheating-school-chatgpt-4f89a552e9093ce2180471b4d4736675 - Education Week, which mentions Cuny only in an Associated Press wire photo caption: https://www.edweek.org/technology/chatgpt-for-teachers-a-boon-a-bust-or-just-meh/2025/11 - OpenAI education newsletter, the company that makes ChatGPT, co-written by Cuny: https://edunewsletter.openai.com/p/how-a-high-school-english-teacher - caseycuny.com, his own site: https://caseycuny.com Sources last verified: September 20, 2026 --- ## The radiologist who saw intimate partner violence coming years before anyone said a word URL: https://goodinprogress.org/stories/bharti-khurana-airs Location: Boston, Massachusetts People: Bharti Khurana Date: July 28, 2026 Dek: Asked whether they felt unsafe in their relationship, patients often said no. Bharti Khurana went looking for the answer in their X-rays instead. - **The person and the place:** Dr. Bharti Khurana is an emergency radiologist at Brigham and Women's Hospital in Boston, and the founder of its Trauma Imaging Research and Innovation Center. She reads the images that decide whether a patient gets admitted or sent home. - **The problem Bharti cared about:** In 2016, on her way out of a trauma hackathon, she heard a nurse practitioner talking about intimate partner violence, and something landed. "For non-accidental trauma in children, we do such an excellent job," she told CBS Boston. "And then, when it comes to adults, we were not doing anything. So, I got motivated." Every patient gets asked whether they feel unsafe in their relationship. Many who are say no. - **The agency moment:** Rather than wait for anyone to disclose, Khurana went after the answer through the pictures she already had. The patterns were sitting on the film: mid-facial fractures, injuries to the orbital bones around the eye, and the particular break in the forearm bone near the pinky that happens when a person raises an arm to shield their face. She and her team built AIRS, a decision-support tool that reads those patterns alongside a patient's clinical history and hands the clinician an assessment while they are still looking at the image. - **What changed, with evidence:** The team published its evaluation in npj Women's Health in March 2026. Working from the records of 841 patients and comparing them against 5,212 others, they measured their strongest model at an AUC of 0.88, a score of how cleanly a model separates two groups rather than a percentage of correct calls, and it stayed between 0.82 and 0.88 across three further groups of patients. At the threshold the team chose, the model reached above 73% sensitivity and specificity. On average it saw the risk more than three years ahead of the day a patient asked for help. By late 2024 the hospital was putting the tool's current iteration to work in its emergency department and at some primary care sites. - **The move worth copying:** She noticed her field did something carefully for one group of patients and almost nothing for another, and treated that gap as hers to work on. She also built for the patient who still says no. AIRS comes with what the team calls "empower guides" and a QR code a patient can scan instead of carrying home a pamphlet that might anger an abuser, and any flag lives in a "safe zone" visible only to the patient and their clinicians, never in the patient portal. The team is clear that the model needs testing in broader populations before wider clinical use, and that it was never built to diagnose anyone. Know someone who fixed the gap everyone else stepped around? Feed the scout at goodinprogress.org. - **Takeaway line:** "The goal is never to force disclosure, but to help clinicians communicate with patients in a supportive way and to connect them with resources and support." — Dr. Bharti Khurana, NIH Research Matters Sources: - npj Women's Health, the team's own paper: https://doi.org/10.1038/s44294-025-00126-3 - NIH Research Matters, the study's funder: https://www.nih.gov/news-events/nih-research-matters/ai-tool-predicts-patients-risk-intimate-partner-violence - CBS Boston: https://www.cbsnews.com/boston/news/boston-doctor-tool-identify-patients-intimate-partner-violence/ Sources last verified: September 13, 2026 --- ## The volunteer who built Japan's disaster filter, then added twenty people to check the AI's work URL: https://goodinprogress.org/stories/kenjiro-murakami-spectee Location: Tokyo, Japan People: Kenjiro Murakami Date: August 3, 2026 Dek: In the wreckage of the 2011 earthquake, Kenjiro Murakami noticed the fastest damage reports were coming from ordinary people's phones, and spent the next decade making them trustworthy enough for a prefecture to reach for. - **The person and the place:** Kenjiro Murakami is the founder and chief executive of Spectee, a company in Tokyo whose stated mission is to make crises visible. It exists because of a trip he took in 2011, when he went to the area hit by the Great East Japan Earthquake as a volunteer. - **The problem Kenjiro cared about:** What he found there was not what he expected. "At that time, I found social media more reliable than mainstream media in gathering the latest information about damage and vital supplies," he told the Government of Japan's KIZUNA magazine. The most useful account of what was actually happening was scattered across strangers' phones, in fragments, mixed in with everything else. - **The agency moment:** He did not wait for the news to get faster. He set out to build the filter instead, and shipped the first version in 2014 with the goal of judging every post automatically. No human in the loop, no bottleneck, just the machine sorting the real reports from the noise. - **What changed, with evidence:** It did not work. The system kept flagging genuine disasters as fake posts, and by around 2015 the company had rebuilt the flow to put people back at the end of it. Roughly twenty people now sit on the verification team, confirming reports by phone and cross-checking them against other data before an alert goes out, and Nikkei xTECH reports that users came to treat the human check as the reassuring part. When the Noto Peninsula earthquake hit Ishikawa Prefecture in January 2024, the prefectural government had already adopted the service, and Nikkei xTECH reports its information was used as material for judging evacuation instructions. With roads cut, the service weighed transport ministry road cameras against driving data from cars that had actually been down those roads, supplied by Toyota. One pharmaceutical company used it to identify its supplier factories inside the disaster zone and had the damage status of all of them confirmed by the next day. There is a catch the Noto quake exposed: when a disaster is bad enough, the posts thin out too. The worse it gets, the quieter the phones. - **The move worth copying:** He had the more impressive version and let it go. Twenty people checking a machine's homework is a worse sentence than full automation, and it is the reason a prefectural government picked the tool up. If something you built is only trustworthy with a person standing next to it, that is not a failure to hide. Know someone who shipped the honest version instead of the impressive one? Feed the scout at goodinprogress.org. - **Takeaway line:** "Disaster response is a race against time. It is crucial to identify and provide the most critical information as quickly as possible so emergency responders can focus on efficient action." — Kenjiro Murakami, Government of Japan KIZUNA magazine Sources: - Nikkei xTECH, Risako Kokuji, February 2024, whose second page is paywalled: https://xtech.nikkei.com/atcl/nxt/column/18/00001/08918/ - ITmedia MONOist, Teruno Ando, June 2026: https://monoist.itmedia.co.jp/mn/articles/2606/03/news047.html - Government of Japan, KIZUNA magazine, Murakami's own words in an interview: https://www.japan.go.jp/kizuna/2025/03/next-gen_disaster_tech.html Sources last verified: September 27, 2026 --- ## The son who built a missing-persons system during the five months he spent searching for his father URL: https://goodinprogress.org/stories/jyrone-parker-project-aziz Location: Louisville and Lexington, Kentucky People: Jyrone Parker Date: August 3, 2026 Dek: Jyronna Parker disappeared from Lexington in October 2025. Across the search, his son Jyrone found that the tool his family needed did not exist, so he built it and then gave it to everyone. - **The person and the place:** Jyrone Parker is an engineer in Kentucky. His father, Jyronna "JD" Parker, was 54, a veteran, and was last reported missing from Lexington on October 17, 2025. "My father was the epitome of strength and patience," Parker told LEX18. "He was a gentle parent. Never struck me but held me to a high standard." - **The problem Jyrone cared about:** A search has no headquarters. Tips arrive by phone and by comment thread, photographs sit in one person's camera roll, sightings get relayed and half-remembered, and the family is the only thing connecting any of it to the police. The formal system did not close the gap either: WKYT reports that a Green Alert, the notice Kentucky issues when a veteran at risk goes missing, was issued for his father in late 2025 and then cancelled by Lexington police after they determined he was not at risk. The search ran on for months. - **The agency moment:** He did not wait until it was over to start. While his father was still missing, Parker built the coordination tool the search did not have, and pointed the narrow, specific version of the technology at the narrow, specific problem in front of him. "I had trained the model based on my dad's voice and his photos and videos, and I had these cameras around the city of Lexington," he told FOX 56. One face and one voice, on cameras placed where he had asked permission first. - **What changed, with evidence:** On March 22, 2026, five months into the search, his father's remains were found on the river shoreline at a park in Clarksville, Indiana. Identification took weeks and came down to dental records. He kept building. The platform is open to anyone now, families, neighbors, and police departments alike, and Parker charges nothing for it. A family can post photographs and video, organize a search party, pin recent sightings, and push notifications and email out to everyone involved at once. There is a community alert built in, which he described to LEX18: "Someone can put up what's called an aziz alert if somebody's loved one has gone missing. And it goes out to everyone in the local community." Standing in his own neighborhood, he put the reach of one alert at a couple hundred people. Nobody has been found through it yet. It is built, it is free, and it is waiting. - **The move worth copying:** He built during the search, not after it, and released it with no success story attached. There is a version of this where he waits for a win before telling anyone, and the families who needed it in the meantime never hear about it. Know someone who built the thing that was missing while they were still in the middle of needing it? Feed the scout at goodinprogress.org. - **Takeaway line:** "In an emergency, situations such as these, every second matters, and also just the facilitation of communication." — Jyrone Parker, FOX 56 Sources: - LEX18, Kayleigh Randle, original interview: https://www.lex18.com/news/covering-kentucky/louisville-man-creates-ai-app-to-find-missing-people-after-fathers-body-found-in-ohio-river - FOX 56 Lexington, RJ Johnson, original interview: https://fox56news.com/news/local/lexington/son-develops-software-after-lexington-veteran-found-dead/ - WKYT News Staff, on the recovery and the cancelled Green Alert: https://www.wkyt.com/2026/04/02/lexington-veterans-remains-found-indiana-park/ Sources last verified: September 20, 2026 --- ## The engineer who could not read the account number on the wall, so he taught his phone to read it for him URL: https://goodinprogress.org/stories/sulaiman-adewale-xara Location: Osogbo, Nigeria People: Sulaiman Adewale Date: August 3, 2026 Dek: Sulaiman Adewale sometimes walked out of a shop without buying anything, because the account number posted on the wall was a blur. The workaround he built for his own eyes now runs inside tens of thousands of Nigerians' chats. - **The person and the place:** Sulaiman Adewale is a software engineer from Osogbo, in southwestern Nigeria. He has been writing code since 2010 and taught himself most of it, not out of principle but because university was not affordable, so he went to technical college instead. "I'm a very dogged person," he told Techpoint Africa. "I could have gone to secondary school like my peers, but I wanted hands-on learning, and I really wanted to learn about computers. From Technical College Osogbo, I proceeded to school myself." - **The problem Sulaiman cared about:** In Nigeria, paying often means reading a bank account number off a sign taped to the wall behind the counter and typing it into your phone. Adewale cannot see well at a distance. "I have issues seeing things that are far," he has said. "Whenever I go to buy stuff at the supermarket, I sometimes have to leave because I wouldn't be able to see the account number. Even when I see it, it's hard looking up and down multiple times because most of the numbers are blurry." The everyday transaction everyone else completed without thinking was, for him, a reason to put the shopping down and go home. - **The agency moment:** Instead of squinting harder, he built something that could do the looking. Point a phone camera at the number on the wall, and the thing he built reads it and sends the money. He and his team put it inside WhatsApp rather than making anyone install anything, which was a deliberate choice about who would be able to use it. "I wanted an easier way that carries everybody along in banking, and if you look at it properly, you will see that WhatsApp is what even the oldest people among us use," he told TechCabal. - **What changed, with evidence:** TechCabal's reporters used it themselves for two weeks. It handled transactions from photographs, voice notes, and typed messages, analysed spending, scheduled payments, remembered past conversations, and saved recipients for next time. It was trained on accented Nigerian speech and works in Pidgin as well as English, with Hausa and Yoruba on the way. By its first anniversary the company reported around 50,000 users and roughly ten billion naira moved, about six million dollars. It is not a bank. It runs on partner banks' licences and holds nobody's money. Early on, a payment partner could not handle the rush and new sign-ups had to be paused. And the thing users keep raising is trust. "Being able to bank via WhatsApp without opening another app is convenient, since it works even on a low network connection," one early user, Babatunde Hassan, told TechCabal. "But I'm worried about how our information is secured, and I'm sure that doubt may also hold other people back." - **The move worth copying:** He did not set out to fix financial inclusion in Nigeria. He set out to buy groceries without giving up. The problem in front of your own face is allowed to be the one you work on, and it is often the one you understand well enough to actually solve. What is the thing you have quietly worked around for years? Know someone who built their own way past it? Feed the scout at goodinprogress.org. - **Takeaway line:** "I wanted an easier way that carries everybody along in banking, and if you look at it properly, you will see that WhatsApp is what even the oldest people among us use." — Sulaiman Adewale, TechCabal Sources: - TechCabal, Abubakar Abdulrasheed, July 2025, who tested the product independently for two weeks: https://techcabal.com/2025/07/04/xara-whatsapp-ai-bot/ - Techpoint Africa, April 2026, a critical feature: https://techpoint.africa/feature/nigerian-xara-founder-xai-offer/ - Technext24, Mubarak Bankole, June 2026, a first-anniversary write-up running on company-supplied figures: https://technext24.com/news/whatsapp-xara-crosses-%e2%82%a610-b-transaction/ Sources last verified: September 27, 2026 --- ## The Afghan refugee who landed in America with no idea where to start, so she started asking URL: https://goodinprogress.org/stories/amina-irc-alma-resettlement Location: United States People: Amina Date: August 3, 2026 Dek: Amina arrived from Afghanistan alone, with immigration paperwork, a housing search, and a job hunt all waiting. She got through the first stretch of it one question at a time. - **The person and the place:** Amina came to the United States from Afghanistan on her own, and started over in a country she had never lived in, with the International Rescue Committee helping her resettle. - **The problem Amina cared about:** Everything arrived at once, and none of it was optional. Immigration filings. A lease to understand. Work to find. A system whose shape you are expected to already know. "I was hopeless," she said. "I didn't know where to start." - **The agency moment:** She started asking anyway. The IRC had put an assistant called Alma on WhatsApp in November 2025, reachable at any hour and answering in Dari and Farsi as well as English, Spanish, and Swahili, and she began putting her real questions to it: green cards, where to find legal help, how housing support works. The asking was the easy part to describe and the hard part to do. Then she did the thing that actually moved her situation, which was to act on the answers. - **What changed, with evidence:** She met with accredited lawyers. She secured rental assistance. She started her immigration process. Alma is not a lawyer, and the IRC says so: it points people toward the right next step, it does not advise them. And a chatbot is only as current as what it knows. Writing in Undark, Mona Hedaya, a research fellow at the Center of Conflict and Humanitarian Studies in Doha, put it plainly: "If the chatbot is unaware of rapid changes to on-the-ground conditions, the result could be information that is both wrong and dangerous." Rules change, offices move, a program closes. Anyone leaning on an assistant like this is leaning on how well someone keeps it fed. - **The move worth copying:** Nobody handed her a map. She treated not knowing where to start as a question with an answer somewhere, rather than a verdict on whether she belonged here, and then she followed the answers into rooms with actual lawyers in them. The asking is the part anyone can copy. Know someone who got through something by asking their way through it? Feed the scout at goodinprogress.org. - **Takeaway line:** "It was like having a knowledgeable assistant in my hand, 24/7. I wasn't alone anymore." — Amina, as quoted by the IRC's Signpost director in Rest of World Sources: - Rest of World, an Ideas opinion essay by Andre Heller, the IRC's own Signpost director, not independent reporting: https://restofworld.org/2026/irc-signpost-humanitarian-ai-refugee-assistance/ - International Rescue Committee press release, the organization describing its own program: https://www.rescue.org/press-release/alma - Undark, an opinion column by researcher Mona Hedaya raising the risk critique: https://undark.org/2026/04/30/opinion-ai-humanitarian-decisions/ Sources last verified: September 13, 2026 --- ## The geophysicist who put cameras where nobody was looking, and started beating the 911 call URL: https://goodinprogress.org/stories/neal-driscoll-alertcalifornia Location: California, USA People: Neal Driscoll, Brent Pascua Date: August 4, 2026 Dek: California's worst fires start in the country where no one lives to smell the smoke, so Neal Driscoll made the empty places watched. - **The person and the place:** Neal Driscoll teaches geology and geophysics at UC San Diego. He also founded and directs ALERTCalifornia, a public safety program based there that has been watching the state since 2017. - **The problem Neal cared about:** A wildfire is easiest to stop in its first minutes. The ones that get away are usually burning where nobody lives, with no neighbor to smell smoke and nobody to call it in until the fire is already big. - **The agency moment:** Rather than wait for more eyes, Driscoll and the ALERTCalifornia team pointed a camera network at the backcountry, and after UC San Diego partnered with CAL FIRE in 2023 they had the feeds watched around the clock for the first thin change in smoke color or density. The cameras sweep a full circle about every two minutes and see 60 miles on a clear day, 120 on a clear night. People stayed in the loop on purpose: analysts confirm each detection, which Driscoll says both holds false alarms down and makes the system more accurate over time. - **What changed, with evidence:** "The AI that's being run on the cameras is actually beating 911 calls," Driscoll told the Associated Press. By the program's own count, ALERTCalifornia informed CAL FIRE's response to roughly 3,600 wildfire incidents in 2025, and in more than half of them fire officials knew about the smoke before a 911 caller reported it. "We can fight fires in the incipient phase before they get too large," Driscoll says. - **Where it led:** The network is past 1,200 cameras now, feeding all 21 emergency command centers in the state, and the same footage does a second job after the fire is out. The cameras show the burned slopes where the roots that held the soil are gone, which is where the next disaster tends to start. - **The move worth copying:** The cameras are the expensive part, and most people cannot build those. The cheap part is the decision: he put watchers on the empty country instead of waiting for it to fill up with people who would notice. - **Takeaway line:** "In many cases, we've started a response before 911 was even called, and in a few cases, we've actually started a response, went there, put the fire out, and never received a 911 call." — Brent Pascua, Cal Fire battalion chief, Associated Press Sources: - Route Fifty, Kaitlyn Levinson, July 2026: https://www.route-fifty.com/public-safety/2026/07/what-california-has-learned-ai-enabled-early-wildfire-detection-system/414566/ - Associated Press, Dorany Pineda and Brittany Peterson, republished by KPBS, May 2026: https://www.kpbs.org/news/science-technology/2026/05/05/states-across-the-wildfire-prone-western-us-are-using-ai-for-early-detection - ALERTCalifornia's own program site: https://alertcalifornia.org/ Sources last verified: August 4, 2026 --- ## Catherine Nakalembe walked Eastern Africa's farms for six years. Then she put the cameras on other people's helmets. URL: https://goodinprogress.org/stories/catherine-nakalembe-helmets-labeling-crops Location: Uganda, Kenya, and across Eastern Africa People: Catherine Nakalembe Date: August 6, 2026 Dek: Satellites could photograph Kenya's farmland but could not tell maize from beans, so a Ugandan geographer, thirteen colleagues, and a network of motorcycle riders went and collected the answer from the roadside. - **The person and the place:** Catherine Nakalembe is a geographer, originally from Uganda, who leads the Africa program for NASA Harvest, the United States space agency's food security and agriculture program. - **The problem Catherine cared about:** Satellites photograph Africa's farmland from roughly 350 miles up. What they could not do was tell her whether the crop below was maize or wheat, and without that, no model could work out how a field would hold up against a heat spike or a long wet season. The pictures were never the hard part. The labels were. - **The agency moment:** So she went and got the labels on foot. Every July and August from 2010 to 2016 she crossed small farms, often no more than three acres each, writing down what was growing where. "I walked so many miles, I was in such great shape," she says of those years, and "It was time-consuming and so hard, but it's what was needed." One person cannot walk a country, so she stopped trying to. She put the cameras on other people's helmets instead, working with institutional partners and a citizen science network of riders who carried GoPros down ordinary roads. - **What changed:** Across Kenya's 2021 and 2022 long-rain seasons, the riders gathered the pictures and Nakalembe and thirteen colleagues built the deep learning pipeline that read them, turning roadside photographs into 4,925 validated crop-type data points. Kenneth Mwangi, Jane Kioko, and Christopher Atsianzale Wakhanala are on that author list beside her. They published the dataset open access in 2025, free for anyone to use, and maps built this way feed yield estimates, which tell a government months ahead whether it is heading for a shortfall. - **Where it led:** She has kept arguing that the tools have to fit the places they are used. Models built elsewhere, she told Rest of World, "often also fail to account for the contexts of the Global South, including high internet costs, limited bandwidth, and a lack of labeled training data." - **The move worth copying:** She walked the farms for six years before she handed anyone a camera. Her question is the one to ask of any model built somewhere else: does it account for where it is actually going? - **Takeaway line:** "So imagine face detection software or Google Street View, but in this case, we're looking for crops, not buildings or faces." — Catherine Nakalembe, Social Science Space Sources: - Scientific Data, the team's own peer-reviewed data descriptor, open access: https://www.nature.com/articles/s41597-025-05762-7 - Social Science Space, her own words in a Quick Insight interview: https://www.socialsciencespace.com/2026/07/catherine-nakalembe-on-geospatial-ai/ - Rest of World: https://restofworld.org/2026/ai-agriculture-local-data/ - Devex: https://www.devex.com/news/how-a-nasa-backed-scientist-uses-satellites-to-help-african-farmers-105010 - Maryland Today, her own university's newsroom: https://today.umd.edu/growing-resilience-from-the-ground-up - Eos, the American Geophysical Union's news outlet, on her own paper: https://eos.org/editor-highlights/how-satellite-data-helped-avoid-hunger-from-drought Sources last verified: August 6, 2026 --- ## Moshe Nussbaum lost his voice to ALS, so he wrote his way back on air URL: https://goodinprogress.org/stories/moshe-nussbaum-ai-voice-return Location: Tel Aviv, Israel People: Moshe Nussbaum, Liron Dahan Date: August 10, 2026 Dek: For forty years he was the voice Israelis heard when something bad happened. When ALS took it, he and his newsroom trained a model on his own archive, and he went back to writing the news himself. - **The person and the place:** Moshe Nussbaum, 71, "Nussi" to generations of viewers, has covered crime and national security for Channel 12 News in Tel Aviv since the channel signed on in 1993. - **The problem Moshe cared about:** ALS came for his speech. He kept reporting as long as he could, then the appearances thinned out and stopped. The job was still in him. The voice that delivered it was going. - **The agency moment:** Forty years of broadcasts were sitting in the station's archive, which meant thousands of hours of him talking. Channel 12's own engineers, led by Liron Dahan, trained a model on it. Nussbaum kept the part of the work the disease could not reach. He reports the story, he writes every word, and the model says it in his cadence. - **What changed:** Channel 12 aired a preview in January 2025 that put his real, halting speech next to the restored voice. Two months later a Haaretz television critic sat down and reviewed an actual Nussbaum report, on a story about the prime minister, the way critics review anyone. He was back on the air, not every night, but genuinely back. He was also the first to name what the tool could do in worse hands, telling AP he worries how easily voices like his could be used to spread falsehoods. And he was plain about its limits: it cannot go live, so the field reporting he loved most is not coming back. - **The move worth copying:** He did not argue about everything he had lost. He found the piece of the job that never needed his voice, the reporting and the writing, and rebuilt the rest around it. If you know someone being quietly written off at work, this is the one to send them. - **Takeaway line:** "The fact that Channel 12 and my news managers are allowing me to reinvent myself anew, that is one of the most important medicines I can get in my fight with this disease." — Moshe Nussbaum, Associated Press Sources: - Associated Press, Melanie Lidman, January 2025: https://apnews.com/article/israel-news-ai-voice-technology-text-speech-als-9a207784f0fe0287a25aa80256c50102 - Ynetnews, Ran Boker: https://www.ynetnews.com/culture/article/h1pzinyukl - Haaretz, Yasmin Levi, TV review, March 2025: https://www.haaretz.co.il/gallery/television/tv-review/2025-03-24/ty-article/.highlight/00000195-c6d5-db42-afb7-c7fded000000 - Ynet Hebrew edition: https://www.ynet.co.il/entertainment/article/hyz85ndljl - The Times of Israel, republishing the AP report: https://www.timesofisrael.com/ai-helps-veteran-israeli-tv-reporter-return-to-the-air-after-disease-took-his-voice/ Sources last verified: August 10, 2026 --- ## The public defender who turned down the AI tools on the market, then had one built for public defense URL: https://goodinprogress.org/stories/jennifer-sellitti-opd-resource-library Location: Trenton, New Jersey People: Jennifer Sellitti, Peter Henderson, Dominik Stammbach Date: August 11, 2026 Dek: Commercial AI could not meet the demands of New Jersey's public defenders, so Jennifer Sellitti's office and a team of Princeton researchers built their own: a plain-language search across 3,000 briefs, 200 internal documents, and 350 directives. - **The person and the place:** Jennifer Sellitti is the New Jersey Public Defender, and the office she runs has been accumulating hard-won knowledge for decades. Briefs, motions, legal research, training materials, the collected instinct of people who have argued these cases before. - **The problem she cared about:** All of it sat scattered across separate offices, servers, and databases. A defender who wanted an argument a colleague had already made well had to go find it first, and that search costs time that belongs to clients. - **The moment they decided not to settle:** The easy path was to buy something off the shelf. Instead her office and researchers from Princeton stress-tested the commercial AI systems against what public defense actually demands, and the answer came back plain: those tools were not designed for the language, the legal reasoning, or the security this work requires. In the office's own announcement, Sellitti named the choice. "When the existing technology couldn't meet the demands of public defense, we didn't lower our standards. We raised them." - **What they built instead:** The OPD Resource Library opened on July 8, 2026. A defender asks a question in plain language, and the library searches 3,000 briefs, 200 internal documents, and 350 directives, returns the passages that matter with a summary for context, and hands back the original PDF with the exact section highlighted so the lawyer can check the work themselves. It runs inside the state's own secured systems, so client files stay where they belong. Princeton's Peter Henderson and Dominik Stammbach built it around how defenders actually search. By Sellitti's own account, gathering the core legal arguments went from hours to seconds. - **The move:** She did not wait for a vendor to make something safe enough for public defenders. She brought in the researchers who study the work and had them build it alongside the people who do it. Know a public servant who got tired of waiting and built the thing themselves? Feed the scout at goodinprogress.org. - **Takeaway line:** "I always say that AI should be making our lawyers more human, not less. If I can automate the mundane, I give them the opportunity to spend more time sitting across the table from a client, and that's not only what we should be doing, it's what lawyers came to this work to do." (Jennifer Sellitti, New Jersey Public Defender, to Law360, via ABA Journal) Sources: - New Jersey Office of the Public Defender, official announcement: https://www.nj.gov/defender/media/press/20260708_NJ_OFFICE_OF_THE_PUBLIC_DEFENDER_LAUNCHES_PD_RESOURCE_LIBRARY.shtml - ABA Journal, Amanda Robert, July 2026: https://www.abajournal.com/news/article/new-jersey-introduces-first-of-its-kind-ai-tool-for-public-defenders - Law360 Pulse, Jake Maher, July 2026: https://www.law360.com/pulse/articles/2502182/new-jersey-public-defender-debuts-novel-ai-research-tool - Princeton Center for Information Technology Policy, the project partner: https://blog.citp.princeton.edu/2026/07/08/princetons-citp-and-the-nj-office-of-the-public-defender-launch-an-ai-powered-resource-library/ Sources last verified: August 11, 2026 --- ## The bank that took a 3,000-person job down to 135, and says it moved them rather than cut them URL: https://goodinprogress.org/stories/dbs-reskilling-singapore Location: Singapore People: Piyush Gupta, Tan Su Shan Date: August 11, 2026 Dek: When AI started absorbing routine work, DBS's outgoing chief executive said the part most bank bosses leave out, that he could no longer see how to create jobs. His successor has spent the time since moving people into different ones, and is equally plain about the roughly 4,000 temporary workers her promise was never going to cover. - **The people and the place:** DBS is Singapore's largest bank, with 39,721 employees at the end of 2025. Piyush Gupta ran it for fifteen years. Tan Su Shan took over in March 2025 and inherited the problem he had just described in public. - **The problem they were facing:** AI is genuinely absorbing routine banking work. The common response is quiet attrition and reassurances vague enough that nobody has to say who is actually affected. - **The moment somebody said it out loud:** In February 2025, at an IT event in India, Gupta gave up the comfortable line. "In my 15 years of being a CEO, for the first time, I'm struggling to create jobs. I'm struggling to say how I will repurpose people to create jobs." In the same breath he said the bank would add 1,000 new AI positions. - **What changed, counted honestly:** Entry-level production engineering, the routine operations work, went from 3,000 people to 135 by July 2026. The bank says those affected were retrained and moved into other tech roles inside the company. Some 13,000 staff have been earmarked for reskilling and more than 10,000 have begun training, tellers are being redefined toward customer relationships and digital servicing, and there is an internal learning hub with more than 10,000 classes in it. Three things sit beside that and belong next to it: roughly 4,000 temporary and contract roles are being phased out through natural attrition, those 8,000 to 9,000 contract workers were never counted in the bank's headcount in the first place, and DBS's own headcount fell by 1,624 during 2025, which the bank attributes mainly to integration work in India and Taiwan rather than to AI. - **The move:** Tan put the choice in plain terms, and then the bank spent eighteen months doing the more expensive version of it. If you are watching a function get automated near you, the question worth borrowing is the one she asked out loud: what happens to the people who have been here twenty years? Know someone whose employer chose the harder path? Feed the scout at goodinprogress.org. - **Takeaway line:** "Other people might say it's easier, just fire everyone and then hire again, but if you do that, then where is the humanistic appreciation of people who've been with you 30 years, 40 years, 20 years?" (Tan Su Shan, CEO, DBS, in Forbes) Sources: - Fortune, Lionel Lim, February 2025: https://fortune.com/asia/2025/02/25/dbs-singapore-to-cut-temp-jobs-ai-piyush-gupta/ - Forbes, Jessica Tan, August 2026: https://www.forbes.com/sites/jessicatan/2026/08/05/dbs-group-ceo-tan-su-shan-leads-southeast-asias-largest-bank-into-the-ai-era/ - Fintech Singapore, reporting on a Bloomberg interview, November 2025: https://fintechnews.sg/121362/ai/dbs-ai-retraining/ - Fintech Singapore, reporting on The Business Times, April 2026: https://fintechnews.sg/130306/ai/singapore-banks-headcount/ Sources last verified: August 11, 2026 --- ## The Nogoyá agronomist who taught a machine to hold 23-plus years of his own judgment URL: https://goodinprogress.org/stories/diego-alvarez-nogoya-agronomist Location: Nogoyá, Entre Ríos, Argentina Date: August 13, 2026 Dek: An Argentine agronomist spent 23-plus years reading his own land, then built an AI system that can answer like he would, long after he's gone. - **The person and the place:** Diego Álvarez is an agronomic engineer who works his family's land in Nogoyá, a farming town in Argentina's Entre Ríos province. - **The problem he cared about:** For more than 23 years he had been reading his own fields: when to plant, how a given plot handles a dry year, which rotation saves it in a wet one. None of it was written down anywhere but his head. - **The agency moment:** So Álvarez started entering it himself, the rainfall logs, yield sheets, and planting records, into ChatGPT. Then, working with a technology partner rather than waiting for an agtech company to build it for him, he trained the system on the judgment calls themselves, the season-by-season reads on this exact ground that only someone who had worked it for decades would know how to make. - **What changed, with evidence:** The system grew into six named AI agents, among them a climate agent and an administration agent, that now handle invoices and receipts he used to process by hand and offer crop recommendations grounded in his own 23-plus years of calls. By his own count, it freed up at least an hour a day, a tally that's his, not an auditor's. - **The move worth copying:** Álvarez didn't wait for an agtech company to solve this for him. He wrote down what only he knew, until a system could answer his own questions back. If you're the one holding years of judgment that has never made it onto paper, that's the version of this worth trying. Know someone carrying expertise like that? Feed the scout at goodinprogress.org. - **Takeaway line:** "El día de mañana, si mis hijos deciden sembrar el campo, pueden entrar, consultar y les va a contestar como si fuera yo" ("The day my children decide to farm the land, they can go in, ask, and it will answer them as if it were me.") — Diego Álvarez, Elonce Sources: - Elonce: https://www.elonce.com/sociedad/un-ingeniero-de-nogoya-creo-una-herramienta-de-inteligencia-artificial-para-dejar-su-experiencia-en-el-campo.htm - Cadena 3 Argentina: https://www.cadena3.com/noticia/la-argentina-posible/un-productor-de-entre-rios-creo-un-sistema-con-ia-para-que-sus-hijos-hereden-23-anos-de-experiencia_579900 Sources last verified: August 13, 2026 --- ## Doug Gorman turned 8,000 strangers' ideas into a plan for his county URL: https://goodinprogress.org/stories/doug-gorman-bowling-green-ai-town-hall Location: Bowling Green, Kentucky Date: August 19, 2026 Dek: Warren County's judge executive replaced a 23-person town hall with an online experiment that let nearly 8,000 residents vote a million times on what Bowling Green, Kentucky should become. - **The person and the place:** Doug Gorman is the Warren County Judge Executive in Bowling Green, Kentucky, a city adding a population the size of itself over the next 25 years. - **The problem he cared about:** a normal town hall on growth issues draws about 23 people, evenly split for and against, nowhere near enough to represent a county about to commit to decisions on housing, schools, and healthcare through 2050. - **The agency moment:** Gorman's office opened an anonymous online idea board called "What Could BG Be?," then brought in Google Jigsaw's Sensemaker, an AI tool built to read thousands of open-ended answers and find where strangers actually agree, as the tool's first large-scale test at government scale. - **What changed, with evidence:** over 33 days, nearly 8,000 residents, about one in ten Bowling Green residents, cast more than a million votes on each other's ideas. Stripped of names and politics, the responses landed on 80 percent agreement across more than half the proposals, and a report that would have taken county staff days or weeks to compile came back in minutes. "What we just conducted was the largest town hall in America," Gorman said. - **The move worth copying:** Gorman didn't wait for a bigger budget or a room big enough to hold everyone. He opened a website, let the county tell him what it wanted, and used AI to actually hear all of them at once. If your town's next planning meeting is heading for the same 23 chairs, point whoever runs it at Bowling Green. - **Takeaway line:** "And what we just conducted was the largest town hall in America." — Doug Gorman, PBS News Hour Sources: - PBS News Hour: https://www.pbs.org/newshour/show/how-a-kentucky-community-is-using-ai-to-help-people-find-common-ground - StateScoop: https://statescoop.com/bowling-green-kentucky-ai-community-planning/ - GovTech: https://www.govtech.com/artificial-intelligence/bowling-green-ky-used-ai-to-drive-public-engagement Sources last verified: August 19, 2026 --- ## The professor who spent decades chasing an unreadable scroll, then handed the finish to a thousand strangers URL: https://goodinprogress.org/stories/brent-seales-herculaneum-scroll Location: Lexington, Kentucky Date: August 20, 2026 Dek: Brent Seales spent decades building software to read Herculaneum's carbonized scrolls without unrolling them, then handed the last mile to a thousand strangers who finally read one all the way through. - **The person and the place:** Brent Seales is a computer scientist at the University of Kentucky in Lexington, where he holds the Stanley and Karen Pigman Chair of Heritage Science. - **The problem he cared about:** the Herculaneum scrolls, carbonized when Mount Vesuvius buried the town in 79 AD, hold philosophical texts nobody has read since. Physically unrolling them destroys them. Centuries of attempts already had, leaving only a compact inner core of some scrolls intact. The texts were, in Seales's words, physically preserved but intellectually inaccessible. - **The agency moment:** Seales spent decades at his EduceLab building software to read the scrolls without opening them: scanning the papyrus, then teaching a model to find the ink no eye could see. In 2023, rather than keep chasing the answer himself, he co-founded the Vesuvius Challenge with Nat Friedman and Daniel Gross, and turned his life's problem into an open contest. "What the challenge allowed us to do was to enlist more than a thousand research teams to work on a problem that would normally have about five people working on it," he said. - **What changed, with evidence:** on June 25, 2026, a team led by papyrologist Federica Nicolardi, working with Seales and the wider Challenge community, announced they had virtually unwrapped and read the entire surviving core of scroll PHerc. 1667, its outer layers destroyed by opening attempts decades and centuries earlier, recovering nearly 1.5 meters of continuous Greek text across 20 columns. It is a philosophical treatise on ethics, likely Stoic, its author still an open question. "This is history," Nicolardi said. "This is completely changing the discipline." - **The move worth copying:** most of the people who read that scroll started as contestants, not classicists. "Most of the Vesuvius Challenge research team first arrived as contestants." If you have spent years alone with a problem, the move Seales made is the one worth trying: open it up before you finish it. The Challenge is still reading, dozens of scrolls remain. Know someone who's stubborn enough to sit with a hard problem, then generous enough to hand it to strangers? Feed the scout at goodinprogress.org. - **Takeaway line:** "Today, we are hearing voices that have been silent for 2,000 years." — Brent Seales, University of Kentucky UKNow Sources: - University of Kentucky UKNow: https://uknow.uky.edu/research/day-herculaneum-scrolls-began-speaking-again - National Geographic: https://www.nationalgeographic.com/history/article/herculaneum-scrolls-mount-vesuvius-ai - Archaeology Magazine: https://archaeology.org/news/2026/06/29/entire-carbonized-herculaneum-scroll-virtually-unwrapped/ - Vesuvius Challenge: https://scrollprize.org/firstscroll - UK Research: https://research.uky.edu/news/grand-prize-discovery-made-2000-year-old-herculaneum-scrolls - arXiv preprint: https://arxiv.org/abs/2606.29085 Sources last verified: August 20, 2026 --- ## Dawn Takeuchi Apuna made Honolulu's building permits work like TurboTax URL: https://goodinprogress.org/stories/dawn-takeuchi-apuna-honolulu-permits Location: Honolulu, Hawai'i Date: August 20, 2026 Dek: Honolulu's top permitting official pointed an AI reviewer at building-permit paperwork so applicants find out what's wrong with their plans before they wait months to hear it, and set a date to make it standard for everyone. - **The person and the place:** Dawn Takeuchi Apuna directs Honolulu's Department of Planning and Permitting, which handles roughly 20,000 building permit applications a year for the city and county. - **The problem Apuna cared about:** Honolulu's permitting system ran on outdated software, and applicants were routinely bounced back for incomplete or noncompliant plans, adding months to projects before anyone told them what to fix. - **The agency moment:** Apuna's department piloted an AI plan-review tool, CivCheck, starting in 2022, built on a platform from Clariti Software, and required it for residential permit applicants starting September 1, 2026. Until then, a temporary incentive program, Priority Review, is getting applicants to try it early. "It's kind of like TurboTax for permitting," Apuna said. "The AI really helps guide applicants through a much better quality application, and then in turn it means less review cycles and faster turnaround." - **What changed, with evidence:** in the department's own comparison, average review time on permits that went through CivCheck fell from 73 days to about 32, and the average number of review cycles dropped from 3.4 to 1.4, a difference DPP says held across an early sample of 19 CivCheck permits against 17 similar ones that skipped it. "We recognize that 19 completed applications is a relatively small sample," DPP spokesperson Davis Pitner said, "however, those results are consistent with what we observed during DPP's earlier six-month CivCheck pilot program." - **The move worth copying:** Apuna didn't wait for a perfect dataset before fixing a broken process. She ran the pilot, let the numbers build, then picked a hard date to make it standard for every residential applicant instead of leaving it optional forever. Know a city office still making people guess what's wrong with their paperwork? Feed the scout at goodinprogress.org. - **Takeaway line:** "It's kind of like TurboTax for permitting." — Dawn Takeuchi Apuna, StateScoop Sources: - StateScoop: https://statescoop.com/honolulu-planning-office-ai-tool-reduce-mistakes/ - Hawai'i Public Radio: https://www.hawaiipublicradio.org/the-conversation/2026-08-12/how-the-ai-tech-required-for-residential-permits-works - Honolulu Star-Advertiser: https://www.staradvertiser.com/2026/07/02/hawaii-news/dpp-launches-fast-track-permit-review-program/ Sources last verified: August 20, 2026 --- ## The legally blind scientist who built himself back into the produce aisle URL: https://goodinprogress.org/stories/john-ross-rizzo Location: New York, NY Date: August 25, 2026 Dek: A physician-scientist who's been legally blind since his teens got tired of being told to just use a delivery app, so he and his lab built the tools that let him shop, ride the subway, and find his way through a building on his own terms. - **The person and the place:** Dr. John-Ross Rizzo is a physician-scientist at NYU Langone Health in New York City who has been legally blind since he was a teenager, from a condition called choroideremia. - **The problem he cared about:** people kept telling him to just use a grocery delivery app instead of shopping himself. "That doesn't support autonomy or agency," he said. "And I have to be honest with you, I like to pick my own damn produce." - **The agency moment:** rather than settle for the workaround, Rizzo turned his own lab's research, on AI mapping systems built for self-driving cars, toward the errands he was tired of outsourcing. Working with fellow researcher Dr. Giles Hamilton-Fletcher and the rest of the VMIL lab he directs, he built Shopscout, a wearable backpack that finds groceries on a shelf with its own camera, Commute Booster, an app that reads subway signs aloud, and Unav, an app that maps a building's inside without cell service. - **What changed, with evidence:** Hamilton-Fletcher's team found that blind and low-vision travelers completed their subway journeys 33 percent faster using Commute Booster. A shopper finds a product on a shelf using Shopscout's own camera, instead of asking a stranger or a delivery algorithm to choose for them. The tools are moving from lab prototypes toward wider use, including future hospital and senior-care settings. - **The move worth copying:** Rizzo didn't start by asking what accessibility technology should look like in the abstract. He started with the specific errand that was still hard for him that week, built it with the people who live the same problem, and let his lab test whether it worked for them too. Know someone stuck working around a problem that could just be solved? Feed the scout at goodinprogress.org. - **Takeaway line:** "I like to pick my own damn produce." — Dr. John-Ross Rizzo, CBS New York Sources: - CBS New York: https://www.cbsnews.com/newyork/news/new-ai-tools-to-help-visually-impaired-people/ - Chelsea News NY: https://www.chelseanewsny.com/features/people/dotty-2024-awardee-dr-john-ross-rizzo-clearing-roadblocks-for-the-visually-impaired-KF3214936 - NYU Grossman School of Medicine: https://med.nyu.edu/faculty/johnross-rizzo Sources last verified: August 25, 2026 --- ## Joshua Tilton built an AI tool that found the heart attacks his own crew was missing URL: https://goodinprogress.org/stories/joshua-tilton-ems-ai Location: United States Date: September 1, 2026 Dek: A rural Ohio paramedic spent his own money on a doctoral AI project, then let it prove him wrong about his own crew's blind spot. - **The person and the place:** Joshua Tilton has worked as a paramedic and EMS instructor for more than 20 years. He's the retired chief clinical officer for the Malta and McConnelsville Fire Department in rural southeast Ohio, where the nearest trauma center is at least an hour away. - **The problem he cared about:** a small volunteer department with no quality-assurance staff had no real way to see what its own crews were missing on calls, or to train for the specific gaps instead of guessing. - **The agency moment:** while defending a doctoral thesis in artificial intelligence at Capitol Technology University, and building the tool at his own expense, Tilton created Artificial Intelligence Quality Assurance, AQUA. It strips patient identifiers from run reports to stay HIPAA-compliant, then runs natural language processing on the narrative notes and vital signs to surface patterns nobody had time to spot by hand. - **What changed, with evidence:** he ran AQUA against months of call data and found crews were under-recognizing atypical heart attack symptoms, the back, shoulder, and stomach pain that don't look like a textbook cardiac call. "As a paramedic myself for 20-plus years, I didn't believe the data," Tilton said. He ran the model again rather than argue with it, then built cardiac training around exactly what it found. Within six months, he said, recognition of those atypical presentations rose sharply. "I know beyond a measure of a doubt that it helped patients," he said. "Then, we saw direct results of that." Those results are Tilton's own account; no independently audited before-and-after figures exist yet. - **The move worth copying:** Tilton didn't need a bigger budget to find his department's blind spot, he needed to trust a number he didn't like enough to check it twice. "I don't know how anyone can afford to not do something like this," he said. Look at what your own records already know before deciding you don't need to ask. - **Takeaway line:** "I know beyond a measure of a doubt that [the system] helped patients." — Joshua Tilton, The Ohio Newsroom / WOUB Sources: - The Ohio Newsroom / WOUB Public Media: https://woub.org/2026/02/12/ohio-fire-department-ai-improve-emergency-care/ - WHIZ / Marquee Broadcasting: https://www.whiznews.com/artificial-intelligence-aiding-first-responders/ Sources last verified: September 1, 2026 --- ## Noella Sudbury wrote the law, then built the app that makes a clean slate fast URL: https://goodinprogress.org/stories/noella-sudbury-rasa-legal Location: United States Date: September 1, 2026 Dek: A former public defender spent years passing Utah's Clean Slate law, then built an AI tool that checks eligibility in three minutes instead of hours. - **The person and the place:** Noella Sudbury, a former Utah public defender and Salt Lake County criminal-justice policy advisor, now founder and CEO of Rasa Legal in Salt Lake City. - **The problem she cared about:** criminal records that keep blocking people from jobs and housing long after they've served their time, in a system where clearing one record the old way took a lawyer 10 to 12 hours. - **The agency moment:** she spent years leading the bipartisan campaign that passed Utah's Clean Slate law in 2019, then turned the same instinct on her own caseload. "Could I build an app? Could I use technology to reduce my time? I just believed it was possible," she said. - **What changed, with evidence:** she built an app that checks expungement eligibility in about three minutes and helps draft the petition, and says it has brought the time a lawyer spends on a case down from 10 to 12 hours to about five. By Rasa's own account, relayed but not independently audited, the company has helped clear the records of 34,000 people across Utah, Arizona, and Pennsylvania since its 2022 launch, with more than 90% of filed petitions approved. One of those people is Virginia Mireles, who spent 17 years incarcerated across Arizona's prison system on five felony convictions before Rasa cleared every one. She now works on Rasa's onboarding team, telling prospective clients her own story. - **The move worth copying:** Sudbury didn't stop at writing a better law, she went back and automated the part that was still slow. "We are decreasing that time every day as we're automating more and more," she said. A rule you fought to change that's still too slow in practice is the second half of the fight. - **Takeaway line:** "Could I build an app? Could I use technology to reduce my time? I just believed it was possible." — Noella Sudbury, NPR Sources: - NPR, All Things Considered: https://www.npr.org/2026/08/03/nx-s1-5892484/ai-legal-tech-jobs-clean-slate - LawNext: https://www.lawnext.com/2023/06/on-lawnext-rasa-legal-founder-noella-sudbury-on-simplifying-criminal-records-expungement.html Sources last verified: September 1, 2026 --- ## The poet who trained an AI on her own poetry, then put it up on the wall at MoMA URL: https://goodinprogress.org/stories/sasha-stiles-a-living-poem-moma Location: New York, NY People: Sasha Stiles, Kris Bones Date: September 1, 2026 Dek: Sasha Stiles fine-tuned a language model on her own poetry until it could answer back in something like her voice, then let a museum audience watch it think out loud for six months. - **The person and the place:** Sasha Stiles is a poet who lives and works near New York City with her husband and studio partner, Kris Bones. She calls herself "a first-generation Kalmyk-American poet and language artist." - **The problem she cared about:** People kept telling her poetry and technology sat on opposite ends of something, one human, one mechanical. She didn't buy the split. "A lot of people think poetry and technology are antithetical, but I find them resonant," she told Scientific American. - **The agency moment:** In 2017 she read about the transformer models starting to power modern AI, and by 2019 she was feeding her own poems into GPT-2 to see what came back. She kept going until she had a real dataset, roughly 200 pages of her own poetry, fine-tuned into a model that knew her own style and vernacular, not just the poetry already sitting in the archives. She named the collaborator Technelegy and published a 2021 collection built from working with it. - **What changed, with evidence:** In 2025, the Museum of Modern Art invited her to fine-tune Technelegy again, this time on metadata from the museum's own text-art collection, and installed the result, A Living Poem, on its lobby wall from September 2025 through the following spring. The piece rewrote itself in verse, animation, and voice every sixty minutes, with a soundscape by Bones. Scientific American profiled the project directly. Not every critic was won over: The Brooklyn Rail's Terry Nguyen called the piece "not poetry, but rhetoric," writing that its output was "no less bland" than the spectacle around it. Stiles has been making her case for years anyway, and keeps building instead of arguing about it. - **The move worth copying:** You don't need a museum wall to try a smaller version. Know someone sitting on a lifetime of their own words, journals, sermons, or field notes? Feed the scout at goodinprogress.org. - **Takeaway line:** "A lot of people think poetry and technology are antithetical, but I find them resonant." — Sasha Stiles, Scientific American Sources: - Sasha Stiles' own site: https://www.sashastiles.com/info - Scientific American: https://www.scientificamerican.com/article/how-a-poet-uses-ai-to-write-and-why-her-work-is-now-at-moma/ - The Brooklyn Rail: https://brooklynrail.org/2026/02/artseen/sasha-stiles-a-living-poem/ - Harvard Magazine: https://www.harvardmagazine.com/arts-culture/harvard-sasha-stiles-poetry-ai-chatgpt Sources last verified: September 1, 2026 --- ## The Cairo student who taught himself to code after losing his sight, then built the app that let his sister see her own safari URL: https://goodinprogress.org/stories/mark-morad-scribeme Location: Cairo, Egypt People: Mark Morad, Bishoy Michael Date: September 1, 2026 Dek: After losing his sight, Mark Morad taught himself to program by screen reader alone and built ScribeMe so no one would have to describe the world to him again. When his sister Karen lost her sight to the same rare disease, she used it to watch a safari unfold. - **The person and the place:** Mark Morad, 20, a student at the American University in Cairo. His sight had been fading since 2018 from a rare, incurable optic nerve condition with no clear timeline. His younger sister Karen's vision failed faster, and by late 2021 they were both completely blind. - **The problem he cared about:** His screen reader could read the words on a page, but went silent the moment a physics or chemistry lesson turned to a chart or a diagram. Explaining every image to him fell to whoever happened to be nearby. - **The agency moment:** "I didn't like that someone always had to explain these documents to me. So I got the idea of creating an app to solve this problem." He had never coded before. He taught himself from YouTube tutorials, working entirely through a screen reader, and had a working version of ScribeMe within five months. - **What changed, with evidence:** Morad's app now points a phone camera, or a pair of Meta smart glasses, at the world and describes out loud what's there, in 20 languages. He runs it today with co-founder Bishoy Michael, and the company reports thousands of users across 140 countries. The clearest evidence of what it actually does: in June 2026, Karen took it on a family safari in Kenya. "I always thought this would be impossible to happen – but it did and I could truly see." - **Where it led:** A gap he built to close for himself, homework nobody could describe to him, is now closing for his sister, on a different continent, over an entirely different kind of picture. He didn't need permission to start, only YouTube and a screen reader. Know someone else stuck describing the world to somebody else, one document at a time? Feed the scout at goodinprogress.org. - **Takeaway line:** "I always thought this would be impossible to happen – but it did and I could truly see." — Karen Morad, quoted in Reuters Sources: - Reuters wire report, carried by Arab News: https://www.arabnews.com/node/2655106/media - AUC News: https://www.aucegypt.edu/news/auc-freshman-develops-scribeme-app-people-visual-impairments - Daily News Egypt: https://www.dailynewsegypt.com/2025/11/11/spanish-innovation-backs-egyptian-tech-scribeme-selected-for-rising-up-in-spain-programme/ Sources last verified: September 1, 2026 --- ## The Kenyan engineer who built Africa's first AI sign-language translator after watching 300 deaf students share one interpreter URL: https://goodinprogress.org/stories/elly-savatia-terp-360-sign-language Location: Nairobi, Kenya People: Elly Savatia Date: September 1, 2026 Dek: Elly Savatia turned motion-captured interpreters into 3D avatars that sign Kenyan Sign Language in real time, working with dozens of deaf and hard-of-hearing Kenyans to teach the system to sign like they do. - **The person and the place:** Elly Savatia, an engineer in Nairobi, Kenya, who traces the drive back to age 10: "how you can use this passion to create a solution to a problem in a community." - **The problem he cared about:** He had watched a class of 300 deaf students share a single interpreter, a gap goodwill alone could not close. - **The agency moment:** In 2023 he founded Signvrse and built Terp 360, motion-capturing real interpreters into 3D avatars that turn speech and text into Kenyan Sign Language in real time. "We like to call it Google Translate for sign language," he says. More than 30 deaf Kenyans have worked with him directly on it, recording the 2,300-plus signs the system now knows. - **What changed, with evidence:** More than 2,000 members of Nairobi's deaf community have used Terp 360's web app, testing with deaf users has pushed its accuracy past 85 percent, and Kenya's National Innovation Agency is now a client. It is honest about what it cannot yet do: English to Kenyan Sign Language only, for now. Leakey Nyabaro, founder of Nairobi's Galaxy Sign Language Training Centre, has said publicly it still has errors and works best in sign-to-voice mode, "which helps the hearing community more than the deaf." Savatia does not dispute the gap: "We're not replacing human interpreters. We're filling the gaps, especially where interpreters aren't available." - **Where it led:** Kelvin Munene, a deaf waiter in Nairobi who has struggled to communicate at hospitals and in job interviews, thinks it will help. No one has yet documented that it has. Know someone stuck bridging a gap like the one three hundred students once shared? Feed the scout at goodinprogress.org. - **Takeaway line:** "We're not replacing human interpreters. We're filling the gaps, especially where interpreters aren't available." — Elly Savatia, CGTN Africa Sources: - Africanews: https://www.africanews.com/2025/08/11/kenyan-startup-signvrse-pioneers-ai-powered-sign-language-translation/ - Royal Academy of Engineering, Africa Prize: https://africaprize.raeng.org.uk/2025-cohort/elly-savatia/ - CGTN Africa: https://newsaf.cgtn.com/news/2026-01-30/When-hands-speak-Terp-360-s-AI-listens-then-translates-1KkFufhhHry/p.html Sources last verified: September 1, 2026 --- ## Vyonne Njeri looked again, and caught the heart murmur a busy day almost missed URL: https://goodinprogress.org/stories/vyonne-njeri-ai-consult-penda-health Location: Nairobi, Kenya Date: September 6, 2026 Dek: A busy clinic day, a software flag, and the moment a clinician trusted her own trained ear over a fast diagnosis. - **The person and the place:** Vyonne Njeri, a registered clinical officer at a Penda Health clinic in Nairobi, Kenya. - **The problem they cared about:** A packed clinic day where a first, fast read is sometimes the only read a patient gets, and the gap between a quick diagnosis and the right one is where something real can slip through. - **The agency moment:** A four-month-old came in with a fever and a stuffy nose. Njeri's first read was a cold. Then AI Consult, a tool reviewing her notes in the background, flagged his elevated heart rate. She didn't take the software's word for it. She picked up her stethoscope, listened again, and heard a murmur consistent with a possible congenital heart defect. - **What changed, with evidence:** "That's something I would have missed on any other day," Njeri told NPR. "That child would have just gone home." Instead, he was referred to a specialist and started on medication. - **The move worth copying:** When something flags a concern, the answer isn't to wave it off, and it isn't to let the flag make the call either. Check it yourself, the way you'd want someone to check on you. - **Takeaway line:** "That's something I would have missed on any other day. That child would have just gone home." — Vyonne Njeri, NPR Sources: - NPR: https://www.npr.org/2026/07/23/g-s1-134929/ai-artificial-intelligence-healthcare - TIME: https://time.com/7304457/ai-prevents-medical-errors-clinics/ Sources last verified: September 6, 2026 --- ## Ivana Drobnjak set out to prove AI doesn't have to waste this much energy, and did URL: https://goodinprogress.org/stories/ivana-drobnjak-ai-energy Location: London, United Kingdom People: Ivana Drobnjak, Maria Perez Ortiz, Hristijan Bosilkovski, Leona Verdadero Date: September 6, 2026 Dek: A UCL professor and her research team measured, rather than assumed, how much power generative AI wastes, and found simple fixes that cut it by up to 90 percent without hurting accuracy. - **The person and the place:** Ivana Drobnjak is a professor at UCL's Computer Science department and a member of the UNESCO Chair in AI, in London. She worked with Maria Perez Ortiz, also at UCL, Hristijan Bosilkovski, a recent UCL graduate, and Leona Verdadero at UNESCO. - **The problem they cared about:** The usual answer to AI's energy problem is more data centers and more power plants. Nobody was asking a simpler question first: how much of that energy is being wasted before it ever reaches a user. - **The agency moment:** So Drobnjak's team stopped assuming and started measuring. They ran a set of original experiments on a real, working model, testing exactly how much energy ordinary engineering choices actually cost. - **What changed, with evidence:** Rounding the model's internal calculations to fewer decimal places cut energy use by up to 44%, while keeping accuracy at 97% or better. Swapping a general-purpose model for one built for a single task cut energy by 15 to 50 times, depending on the job. Cutting response length in half saved another 54%. Stacked together, the changes cut energy use by 90% for repetitive tasks and 75% for complex ones. UCL's own estimate: one day of those savings could power 30,000 to 34,000 UK households. - **The move worth copying:** Most people can't rerun a university's energy experiments. But the underlying habit travels: before reaching for the biggest, most general model for a job done the same way every time, ask whether something smaller, measured against that exact job, would do it for a fraction of the power. Know a researcher measuring what everyone else just assumes? Feed the scout at goodinprogress.org. - **Takeaway line:** "Our research shows that there are relatively simple steps we can take to drastically reduce the energy and resource demands of generative AI, without sacrificing accuracy and without inventing entirely new solutions." — Ivana Drobnjak, UCL News Sources: - UCL News: https://www.ucl.ac.uk/news/2025/jul/practical-changes-could-reduce-ai-energy-demand-90 - TechRepublic: https://www.techrepublic.com/article/news-ai-energy-sustainability-unesco-report/ - TVBEurope: https://www.tvbeurope.com/sustainability/unesco-and-ucl-showcase-90-per-cent-energy-reduction-in-large-language-models - UNESCO: https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90 Sources last verified: September 6, 2026 --- ## The school district CIO who stopped buying software and taught his team to build it URL: https://goodinprogress.org/stories/kris-hagel-vibe-coding Location: Gig Harbor, Washington People: Kris Hagel, James Cantonwine Date: September 6, 2026 Dek: When Peninsula School District's tech chief got tired of paying vendors for simple software, he picked up Claude Code, taught his colleagues to do the same, and the district started canceling contracts instead of renewing them. - **The person and the place:** Kris Hagel runs technology for Peninsula School District, a public school system in Gig Harbor, Washington. He has spent more than two decades in K-12 tech, most of it approving other people's software. - **The problem they cared about:** Every year brought another vendor quote for a tool the district needed: workflow forms, lesson-planning dashboards, and a digital signature system. The prices kept climbing, and building software in-house never looked like a real option for a school IT department. - **The agency moment:** Hagel and Peninsula's research director, James Cantonwine, started prompting Claude Code to write the tools themselves instead. Cantonwine, a former middle school science teacher with no formal coding background, built a lesson-feedback app called LessonLens without learning to code the traditional way. Hagel adapted an open-source e-signature tool the same way. Soon, several other administrators had picked up their own Claude Code subscriptions and started building. - **What changed, with evidence:** By spring 2026, Peninsula had working, homegrown replacements for several paid tools, running on the district's own AI Studio system. One strategic-planning app that would have cost $30,000 to $40,000 from a vendor was built in a few hours instead. The district projects it will save $220,000 to $250,000 by the 2026-27 school year in canceled contracts. Beyond LessonLens, the team has built tools for accounting, human resources, and other back-office work that used to mean writing another check to a vendor. - **The move worth copying:** The habit spreading through Peninsula's office isn't one specific app, it's opening Claude Code before opening a vendor's pricing page. "Just try and see what you can build," Hagel said. If your team has been paying for something a clear enough prompt could handle, try building the small version yourself first. - **Takeaway line:** "Just try and see what you can build." — Kris Hagel, K-12 Dive Sources: - Education Week: https://www.edweek.org/technology/a-district-expects-to-save-200k-from-ai-powered-vibe-coding-heres-how/2026/05 - K-12 Dive: https://www.k12dive.com/news/vibe-coding-helped-a-washington-district-save-250k-in-ed-tech-costs/816993/ - Kris Hagel's own repository: https://github.com/psd401/psd-claude-plugins Sources last verified: September 6, 2026 --- ## Coleman wanted a Rosetta Stone for Owens Valley Paiute. There wasn't one, so Coleman built one. URL: https://goodinprogress.org/stories/jared-coleman-kubishi-paiute Location: Owens Valley, California People: Jared Coleman Date: September 7, 2026 Dek: A computer scientist teaches a language model to speak the endangered tongue of the Big Pine Paiute Tribe, one rule and one elder's sentence at a time. - **The person and the place:** Jared Coleman is a member of the Big Pine Paiute Tribe of the Owens Valley, California, now an assistant professor of computer science at Loyola Marymount University in Los Angeles. - **The problem they cared about:** Five years ago Coleman found an old archive of Owens Valley Paiute: a dictionary and recordings of Coleman's elders, including Coleman's own great-great-great-grandfather's voice. "I had no idea what he was saying," Coleman told WHYY. The archive was enough to learn from by hand, but not enough to train an AI model the usual way. Google Translate-style systems learn from "millions and millions and millions of sentences," Coleman said. "You can't do that for endangered languages, we just don't have that many sentences." - **The agency moment:** In 2023, finishing a PhD at USC with advisor Bhaskar Krishnamachari and researcher Khalil Iskarous, Coleman built something different: a system that learns a language's grammar rules directly instead of waiting for a training set that will never exist. Then Coleman brought in native speakers and elders to train the system by hand, the way you'd teach a person. "So I taught it," Coleman told WHYY. "I said this is the word for water. This is the word for drink." - **What changed, with evidence:** The result, Kubishi (Paiute for "brain"), lets people look up Paiute words, build new sentences, and translate the language. It's publicly available online, and Coleman's Kubishi Research Group keeps expanding it. It is also built not to bluff: "When you're talking with ChatGPT, you should never assume that it's accurate because it just makes stuff up all the time," Coleman told WHYY, which reported that Kubishi "doesn't spit out wrong translations when it doesn't know the answer." CNN independently confirmed the project the following year: the tools are intended "first and foremost" to help Coleman's own community, CNN reported, with hope that visitors to the Owens Valley will take an interest in the language too. - **The move worth copying:** Coleman didn't wait for a company with enough training data to get around to Owens Valley Paiute. When there wasn't enough data to teach a machine the ordinary way, Coleman taught it directly, one rule and one elder's sentence at a time. Know a language running low on data, and someone patient enough to teach it right? Feed the scout at goodinprogress.org. - **Takeaway line:** "You can't do that for endangered languages, we just don't have that many sentences." — Jared Coleman, WHYY Sources: - WHYY: https://whyy.org/segments/endangered-language-speakers-get-creative-with-ai/ - CNN Business: https://www.cnn.com/2025/09/02/tech/ai-endangered-language-preservation - Kubishi Research Group: https://research.kubishi.com Sources last verified: September 7, 2026 --- ## Ivana Feldfeber built the tool that counts what the court files left out URL: https://goodinprogress.org/stories/ivana-feldfeber-aymurai-court-data Location: Buenos Aires, Argentina Date: September 7, 2026 Dek: Argentina had almost no data on violence against women and LGBTI people, so a gender-data observatory and one criminal court built software that hides the victims' names and turns what is left into numbers anyone can use. - **The person and the place:** Ivana Feldfeber is CEO and co-founder of DataGénero, a gender-data observatory in Argentina. Another co-founder, Yasmín Belén Quiroga, is the secretary of Criminal, Contravention and Minor Offenses Court No. 10 in Buenos Aires, and studied data analysis on top of law. The two ends of this story were already connected by a person who worked at both. - **The problem they cared about:** "We found that there was no data on the violence and inequalities that many Latin American women and LGBTI people live with on a daily basis," Feldfeber told KTH. "Often, there was not even data on the gender of the victim of the violence." Two official sources existed, one office inside the Supreme Court and one national registry, and the registry's reports stopped at 2018. - **The agency moment:** Court No. 10 had already spent a decade publishing its decisions, at first by scanning rulings with the names painted out by hand with correction fluid. DataGénero, working with specialists in Mexico and Sweden under a call from the A Plus Alliance for Inclusive Algorithms, built AymurAI to do that part properly: it reads a digitized ruling, swaps out the names and identifying details, and lifts the rest into structured fields: the kind of violence alleged, the gender of the people involved, how the accused was related to the person reporting, where it happened, and how the court ruled. Nothing is automatic. "Es importante entender que la IA nunca debería resolver sola de manera autónoma", Quiroga and Feldfeber told Infobrisas. A person validates every result, and the software installs on the court's own servers so the files never travel to anyone's cloud. - **What changed, with evidence:** Judge Pablo Casas, who runs Court No. 10, says the anonymizing that used to take ten minutes now takes three, and counts the difference as time returned to advising the people in front of the bench. The opened data has become something researchers can work with: analyses run by court staff and by students at the Escuela Argentina de Datos found that in 86% of the court's 2025 gender-violence resolutions, the accused was a partner, an ex-partner, or a family member of the woman who reported. The Buenos Aires judiciary's AI standardization committee formally approved AymurAI on September 24, 2025. It is free, it installs on local servers, and it now anonymizes in Spanish across more than 30 categories, in courts from Chaco to Tucumán and in a prosecutor's office in Costa Rica. - **The move worth copying:** The tool is the second half of this story. The first half is that a single court decided its own records belonged to the public and published them in a form other people could actually use, which is what left something to count. Know a clerk, a registrar, or a records officer sitting on data nobody outside can read? That is where this starts. Feed the scout at goodinprogress.org. - **Takeaway line:** "Es importante entender que la IA nunca debería resolver sola de manera autónoma, siempre tiene que haber una persona que la utilice y valide lo que la IA da como resultado" (It is important to understand that AI should never decide alone, autonomously. There always has to be a person using it and validating what the AI returns.) — Yasmín Quiroga and Ivana Feldfeber, Infobrisas Sources: - Infobrisas: https://www.infobrisas.com/noticias/2023/03/06/53228-como-es-aymurai-la-inteligencia-artificial-que-recolecta-datos-de-violencia-de-genero-en-argentina - Revista Acción: https://accion.coop/2026/08/09/pais/sociedad/inteligencia-artificial-feminismo-tecnofeminismo/ - Feminacida: https://www.feminacida.com.ar/justicia-abierta-caba - KTH Royal Institute of Technology: https://www.kth.se/en/eecs/nyheter/nar-data-gor-skillnad-sa-hjalper-aymurai-rattsvasendet-i-latinamerika-1.1406344 - Consejo de la Magistratura CABA: https://consejo.jusbaires.gob.ar/comite-de-estandarizacion-en-inteligencia-artificial-ia/proyectos-ia/aymurai/ - DataGénero: https://datagenero.substack.com/p/spoiler-con-aymurai-no-nos-quedamos Sources last verified: September 7, 2026 --- ## The fifth-grader who started building a wildfire detector, and kept going until a city bought it URL: https://goodinprogress.org/stories/ryan-honary-sensory-ai Location: Irvine, California Date: September 14, 2026 Dek: Ryan Honary started building wildfire sensors in fifth grade, and in July 2026 the City of Irvine signed on for the first paid deployment, across the Irvine Open Space Preserve. - **The person and the place:** Ryan Honary grew up in Newport Beach, California, and spent high school building wildfire sensors. This fall Honary heads to Stanford University on a Wildfire Fellowship, still directly involved in SensoRy AI, the company Honary founded while in high school. - **The problem they cared about:** In 2018 the Camp Fire burned more than 150,000 acres of Butte County and killed 85 people. Honary was in fifth grade, and started building. - **The agency moment:** Honary did not wait to be old enough. The work started as an elementary school student, looking for a way to catch an ignition in its first minutes, and it kept going: a grant, then a company founded while Honary was still in high school. Along the way it collected an Office of Naval Research Naval Science Award, and in 2019 the grand prize at the Ignite Innovation Student Challenge. - **What changed, with evidence:** The City of Irvine and the Orange County Fire Authority put test sensors in the ground with SensoRy AI last year. The city says they proved "highly effective at detecting wildfire ignitions and providing data on fire risk," and that in one deployment the network "detected and reported an ignition from more than a mile away before it was identified by residents or any camera systems." In July 2026 Irvine signed on for a large-scale network across the Irvine Open Space Preserve, funded by the City of Irvine, the Irvine Ranch Conservancy, and the Regional Forest and Fire Capacity Program at the California Department of Conservation. Michael O'Connell, who leads the conservancy, says it has supported Honary and SensoRy AI since the very beginning. Installation begins in early 2027. Honary told KTLA it is the first paid project. - **The move worth copying:** The sensors are the impressive part. The move worth copying is quieter: Honary treated a fifth-grade science project as unfinished rather than graded, and was still working on it eight years later. Know someone still building the thing they started years ago? Feed the scout at goodinprogress.org. - **Takeaway line:** "With greater than 1,200 feet line of sight and around 300 feet of non-line of sight, we're able to detect fire that's smaller than one foot." — Ryan Honary, NBC Los Angeles Sources: - KTLA: https://ktla.com/news/orange-county/southern-california-city-partners-with-teen-founded-ai-wildfire-detection-service/ - NBC Los Angeles: https://www.nbclosangeles.com/news/local/orange-county-teen-creates-network-of-sensors-to-detect-wildfires/3628276/ - City of Irvine press release, 2026-07-27: https://cityofirvine.gov/news-media/news-article/city-irvine-and-sensory-ai-launch-wildfire-detection-and-risk-assessment Sources last verified: September 14, 2026 --- ## The researcher who built a sepsis warning for the hours before anyone suspects it URL: https://goodinprogress.org/stories/suchi-saria-trews-sepsis Location: Baltimore, Maryland, United States Date: September 20, 2026 Dek: After losing a nephew to sepsis in 2017, Suchi Saria spent years turning a Johns Hopkins lab's early warning system into the first continuous sepsis monitor the FDA has cleared. - **The person and the place:** Suchi Saria teaches at Johns Hopkins University in Baltimore and directs its AI & Healthcare Lab. - **The problem they cared about:** Sepsis is one of the leading causes of death in American hospitals, and it is easy to miss, because fever and confusion belong to plenty of other conditions. By the time a clinician suspects it, the clock has usually been running for hours, sometimes days. Saria lost a nephew to it in 2017. - **The agency moment:** Every other sepsis tool the FDA had authorized waited for a doctor to get suspicious first. Saria's bet was that the suspicion is the delay. Working with a team at Johns Hopkins, Saria built the Targeted Real-Time Early Warning System to read what a hospital already has in its records, the complaint someone arrived with, lab results, vital signs, procedures, and medications, and to raise a flag before anyone on the floor is worried. - **What changed, with evidence:** In 2022 the team published what happened when it ran for real. The Nature Medicine paper carries twenty names, Saria's last: 590,736 patients watched across five hospitals, 6,877 of them flagged with sepsis before antibiotics had started. Among that group, the patients whose alert a provider confirmed inside three hours had an 18.7% adjusted relative reduction in in-hospital mortality, plus less organ failure and shorter stays, compared with the patients whose alert nobody confirmed in time. The software did not manage that by itself. It counted when somebody answered it. In May 2026 the FDA cleared the version Saria's company, Bayesian Health, now sells, the first continuous AI sepsis monitor to get there, and it is running at Cleveland Clinic, MemorialCare, and the University of Rochester. - **The move worth copying:** You are not going to build this one at home, and you do not need to. The part that travels is the question Saria asked before any of the modeling started: what is this process waiting for that it does not actually have to wait for? A suspicion. A form. A Monday. Know someone who stopped waiting for one of those? Feed the scout at goodinprogress.org. - **Takeaway line:** "Sepsis has been the focus of my work for more than a decade — a direction set, in part, by losing someone I loved to it." — Suchi Saria, Healthcare Dive Sources: - Healthcare Dive: https://www.healthcaredive.com/news/bayesian-health-gets-fda-nod-for-ai-sepsis-detection-tool/820107/ - MedCity News: https://medcitynews.com/2026/05/bayesian-sepsis-ai/ - Nature Medicine: https://www.nature.com/articles/s41591-022-01894-0 - Crossref record for the study: https://api.crossref.org/works/10.1038/s41591-022-01894-0 - Johns Hopkins Hub: https://hub.jhu.edu/2026/05/12/fda-approves-early-warning-system-for-sepsis/ Sources last verified: September 20, 2026 --- ## The Eatons counted seven healthy days in Olivia's year, then started a folder called Genome Detectives URL: https://goodinprogress.org/stories/hilary-eaton-genome-detectives Location: Winchester, MA, USA People: Hilary Eaton, Matt Eaton Date: September 21, 2026 Dek: Hilary and Matt Eaton are scientists, but they are not doctors and not geneticists. Over about eighteen months of late nights, the two of them worked through their own family's genome with ChatGPT, reading every paper it cited before they believed any of it. - **The person and the place:** Hilary Eaton and Matt Eaton live in Winchester, Massachusetts, with their daughters Fiona and Olivia. Hilary Eaton has a PhD in molecular cancer biology and did a postdoctoral fellowship at Harvard Medical School. Matt Eaton is a computational biologist who did postdoctoral research at MIT. Neither of them is a physician. - **The problem the Eatons cared about:** Olivia had been sick since infancy. Colds that ran for weeks, scrapes that would not close, a newborn hearing screening failed, and not one night of unbroken sleep. A connective tissue disorder accounted for some of the early physical delays and none of the illness. The parents kept a calendar and marked every day without a respiratory or skin infection. Adding them up, Eaton told the Globe, "it was like seven days out of the year that she was healthy." At the worst of it, PEOPLE reported, the family was carrying more than 150 doctor's appointments a year across 49 specialists. - **The agency moment:** In January 2025, a lab result landed on Eaton's phone in the middle of a work meeting. Olivia's antibody levels were far below where a healthy child's should be, and the doctor was not reachable that minute. "I'm just going to shove everything into it," Eaton recalled deciding, and typed the symptoms and the test results into ChatGPT, asking what the possible diagnoses were. What came back was not a diagnosis, and Eaton has been consistent about that in both interviews. It was a place to start. "It just felt like I could have some iota of agency or control back," Eaton told the Globe, "that I could start having a conversation where I'm not just sort of the passenger [on] a roller coaster that is totally off the rails." - **What changed, with evidence:** In 2023 the Eatons had paid about $4,500 to sequence all four family genomes. Now they worked it after bedtime, saving what they found in a shared folder labeled "Genome Detectives." They set the rules before they started: they told the model what their training was, told it how they wanted it to work, and required a citation for everything, because they intended to read the papers themselves. That last rule earned its keep. When the model flagged a mutation whose literature said those children die between seven and ten, they lost two hours to terror before the real culprit turned up, which was errors in the sequencing data rather than anything in Olivia. The finding that held was a different mutation, one that might explain why a scrape took so long to close and why a Band-Aid could take skin with it. Doctors ordered more genetic testing on the strength of it, and the testing pointed to Okur-Chung neurodevelopmental syndrome, estimated to affect about 1 in 100,000 people. Dr. Wendy Chung, who co-discovered the syndrome and is chief of pediatrics at Boston Children's Hospital, treats Olivia and puts the condition on all of it: the tools work when they are a partnership with medical providers. Olivia started monthly antibody infusions. "She's missed three days of school," Eaton said. "That's it." Later, preparing for a neurology appointment, the chatbot suggested asking about an EEG. Dr. John McLaren agreed it was worth "sort of 'peeking under the hood,'" the test found abnormal brain activity, and on medication Olivia slept through the night, which had not happened once since birth. - **The move worth copying:** Eaton never asked the machine for the answer. "ChatGPT told me where to look," is how PEOPLE has it, and then Eaton went and looked, and took what turned up to people with medical degrees. The sequencing is out of reach for most of us. The rest of it is not: before your next appointment, try asking for the questions you should be asking, ask for the citations, read them, and walk in with a list instead of a worry. - **Takeaway line:** "This is us just amateur spelunking, basically, in our daughter's DNA." — Hilary Eaton, to The Boston Globe Sources: - The Boston Globe, Allyson Chiu: https://www.bostonglobe.com/2026/09/12/metro/ai-chatgpt-health-rare-disease/ - PEOPLE, Lily Brown: https://www.aol.com/articles/150-doctor-appointments-mom-says-191511000.html Sources last verified: September 21, 2026 --- ## The blind runner who helped build the glasses that called his race URL: https://goodinprogress.org/stories/thomas-panek-ai-glasses-nyc-half Location: New York City Date: October 1, 2026 Dek: Thomas Panek had already finished the NYC Half with the help of guide dogs. In March, Panek ran it again with AI glasses calling the turns, a safety runner alongside, and a coach watching from Colorado. - **The person and the place:** Thomas Panek noticed something was wrong young. "I was doing things like everybody else does, riding a bike or running, and I'd break my arms, and bump into things," Panek told amNewYork. "People just thought I was clumsy, and what was really going on was that I was experiencing vision loss." Legally blind by the early twenties, Panek went to Lighthouse Guild in New York at 26 for a first cane and then a guide dog, and now runs the organization as president and chief executive. - **The problem they cared about:** Running meant borrowing somebody else's eyes. Panek had pushed at that before, finishing the 2019 NYC Half with the help of guide dogs. The open question was whether the course itself, the turns, the water stops, the crowd, could reach a blind runner without a person out front reading it aloud. - **The agency moment:** Panek worked directly with Meta to customize a pair of AI glasses for a race course, and was clear about how that work should go. It matters, Panek told CBS News New York, that the devices are "developed by people who are blind, for people who are blind." "That's the first time that kind of approach has been taking place." - **What changed, with evidence:** On March 15, 2026, Panek ran the 13.1 miles of the United Airlines NYC Half from Brooklyn to Central Park with the glasses calling the course, the water stops, the turns, and the landmarks. Two other people were in it. Jed Laskowitz ran alongside as a safety runner for Achilles, and the ultra-marathoner Scott Jurek watched the race through Panek's sight line from home in Colorado and coached into it. Panek's own accounting is precise: "I had a safety runner, Jed Laskowitz, who ran on behalf of Achilles, but at the end of the day, that was the first time I had run semi-independently since I was a boy." The stretch Panek called the most emotional was hearing the glasses describe the people watching and what shirts they were wearing. - **The move worth copying:** Michael Cush, who was born legally blind and is chief program officer at Visions in New York, teaches other blind and low-vision New Yorkers to use tools like these, and keeps the claim the right size. A cane or a guide dog is not being replaced. "The goal is more choices, more independence." The portable part here is not the glasses. Those took a corporate engineering team to build. Panek emphasizes involving the people who will use a thing in making it. Know someone who got into that room and changed what came out of it? Feed the scout at goodinprogress.org. - **Takeaway line:** "I always want to do it on my own. I've been able to achieve that dream." — Thomas Panek, CBS News New York Sources: - CBS News New York: https://www.cbsnews.com/newyork/news/ai-powered-technology-helps-blind-people-gain-independence/ - WBUR Here & Now: https://www.wbur.org/hereandnow/2026/04/01/thomas-panek-nyc-marathon - amNewYork: https://www.amny.com/news/thomas-panek-nyc-half-marathon-lighthouse-guild/ - CBS News New York: https://www.cbsnews.com/newyork/news/blind-man-runs-nyc-half-with-guide-dogs/ Sources last verified: October 1, 2026 --- ## Andrew Ngui put sixty people and a drone over the same neighborhood, just to see which was faster URL: https://goodinprogress.org/stories/andrew-ngui-kansas-city-drone-assessment Location: Kansas City, MO, USA People: Andrew Ngui, Joshua Jeffrey, Geoff Hinkle Date: October 1, 2026 Dek: After a storm, Kansas City counts damaged homes on paper, door to door, before federal aid can move. Andrew Ngui spent one March morning testing whether drones and an AI model could do that count, with about sixty ground assessors working the same blocks as the control. - **The person and the place:** Andrew Ngui is chief digital officer for Kansas City, Missouri, which GovTech measures at about 319 square miles of ground to check after a bad storm. - **The problem Andrew cared about:** Federal aid does not move until the city files a preliminary damage assessment with FEMA, and that assessment has been a paper exercise. The city trains staff and volunteers in disaster management, then sends them out on foot to work the damaged blocks, one house at a time. "Versus always being caught like a deer in the headlights," Ngui told KSHB. "That's really the heart of it." - **The agency moment:** On March 23, 2026, Ngui's office ran the question as a simulation, on a day when no storm had hit and nothing was at stake. The city flew its own drones over one neighborhood, and about sixty people worked the same ground the old way at the same time, so there would be two results to hold up against each other. Joshua Jeffrey, program manager for Bellwether, the Google X team whose model read the images, described the handoff plainly: "The city is flying their own drones, they'll capture images of homes and then we access those images to see how much damage was done." Kansas City's emergency management coordinator, Geoff Hinkle, brought in city fire and police, Platte County, and the National Weather Service. - **What changed, with evidence:** Ngui's team pointed the model at one narrow job, sorting what the drones saw into FEMA's damage scale of one to five. By Bellwether's own count, processing that imagery took about 23 minutes in this scenario, and the model landed within one number on the scale 96 percent of the time. Those are the tool team's figures rather than an outside measurement, and Google paid for the project, which is worth saying out loud. What nobody has to take on faith is the shape of the morning: one neighborhood, two methods, sixty people on the ground, and a city that now has its own look at what it would be choosing between. - **The move worth copying:** Ngui says the exercise could help create a playbook for other cities. Know someone at a city hall who tested the new way against the old way before anybody's roof was actually gone? Feed the scout at goodinprogress.org. - **Takeaway line:** "Because it's not just about the damage; it's about really restoring people's lives as quickly as possible to the status quo." — Andrew Ngui, to Government Technology Sources: - Government Technology / GovTech, Julia Edinger: https://www.govtech.com/artificial-intelligence/kansas-city-mo-turns-to-ai-to-improve-disaster-response - KSHB 41 Kansas City: https://www.kshb.com/news/local-news/missouri/clay-county/city-of-kansas-city-mo-first-to-use-new-ai-and-drone-technology-to-assess-severe-weather-damage-to-homes Sources last verified: October 1, 2026 --- ## The software engineer who made his town's budget readable, then told his neighbors to audit him URL: https://goodinprogress.org/stories/andrew-baber-marblehead-budget-data Location: Marblehead, Massachusetts Date: October 1, 2026 Dek: Andrew Baber pointed an AI at 3,000 pages of Marblehead's own financial documents, published the charts before the override vote, and put the whole repo online so anyone could go looking for the bias. - **The person and the place:** Andrew Baber moved to Marblehead, Massachusetts in 2017 and stayed out of town arguments for years. A software engineer by profession, Baber holds no town office, and the site's own disclosure page says so before it says anything else: "Made by Andrew Baber, a Marblehead homeowner and parent. Not on the Select Board, the Finance Committee, or any other town body, and not affiliated with any campaign or advocacy group." - **The problem they cared about:** In the spring of 2026 Marblehead faced three competing override questions, at $9 million, $12 million, and $15 million, and the debate had outrun the arithmetic. "I found it really difficult to have actual discussions that weren't just emotional," Baber told Itemlive. "People would say stuff that felt like there were some facts behind it, but it was really hard to actually dig into those facts." - **The agency moment:** The numbers were already public, sitting in 39 primary-source PDFs running past 3,000 pages. Baber started charting them and posting the screenshots to Facebook, then built a page so residents could look for themselves. Knowing exactly how software gets made, Baber chose to direct Claude Code rather than write it. "I would say maybe not 100%, but certainly north of 99%," Baber said of the division of work. "I have not written a single line of code." - **What changed, with evidence:** In roughly five days the page became marbleheaddata.org, logging more than 700 code updates and over 500 feature changes, with charts on school staffing, insurance trends, peer-town comparisons, and a calculator for what each tier would cost a household. "People found it really helpful," Baber told WBZ NewsRadio. Baber publicly supported the override, and says plainly that some residents may read the site through that lens: "I tried really hard to be extremely center of the road. Here's the data. Here are the facts. Make of it what you will." On June 9 the town approved the $15 million question, 4,278 to 3,594, on 8,092 ballots, the heaviest turnout at an annual town election in the Marblehead Independent's records since 2006. In July, Itemlive reported the site had "helped many residents better understand the town's finances." - **The move worth copying:** Baber is candid that the machine makes mistakes and that not every number has been checked back to its source document by hand. So the whole thing sits on GitHub with a prompt published for anyone who wants to come after it: "audit this site. find the bias. follow the money. be skeptical." Your town's budget documents are almost certainly online too, and almost certainly unreadable. Start with one chart, then hand it to somebody who disagrees with you. - **Takeaway line:** "I want this site to be a valuable tool for anyone coming to it. Whether you're a hard yes, hard no … or somewhere in the middle." — Andrew Baber, Marblehead Current Sources: - Marblehead Current: https://marbleheadcurrent.org/2026/04/21/new-resident-built-website-uses-ai-to-make-sense-of-towns-budget-ahead-of-override-vote/ - Itemlive: https://itemlive.com/2026/07/19/marblehead-budget-website-grows-beyond-override/ - WBZ NewsRadio 1030: https://wbznewsradio.iheart.com/content/marblehead-man-creates-website-to-share-data-on-town-finances/ - Marblehead Independent: https://www.marbleheadindependent.com/marblehead-voters-approve-15m-override-ending-more-than-two-decades-of-resistance/ - Marblehead Budget Data: https://marbleheaddata.org/about Sources last verified: October 1, 2026 --- ## The Nairobi engineer who built a robotic hand to sign chemistry to deaf students URL: https://goodinprogress.org/stories/norah-kimathi-zerobionic-sign-language Location: Nairobi, Kenya People: Norah Kimathi Date: October 2, 2026 Dek: It took Norah Kimathi and the ZeroBionic team seven to nine hours to get one chemical equation across to students who could not hear it, so they built a hand that signs instead, and asked deaf teachers to record the vocabulary it needed. - **The person and the place:** Norah Kimathi grew up in Nairobi West, Kenya, where floods were a regular fact of life. At 15, with no formal technical training, Kimathi built a small phone out of Lego bricks. "I started tinkering with devices that could help me and my family go about our daily activities in a normal way whenever such incidents would occur," Kimathi told the Daily Nation. Strathmore University came later, and a degree there in Informatics and Computer Science, in a country where women were 30.1 percent of ICT graduates in 2022, the most recent government count. - **The problem they cared about:** Mentoring younger students in STEM through Young Scientists Kenya, Kimathi and the team kept walking into classrooms where some of the students could not hear them. "We happened to encounter some kids who could not hear, and some actually couldn't speak, so we had to teach using images," Kimathi told the Daily Nation. One chemical equation ate a working day: "It took us about seven to nine hours of frustration." The shortage is structural, not local. Roughly 300,000 children in Kenya have a hearing impairment and only about 20,000 were enrolled in school as of 2024, according to the Kenya Society for Deaf Children, and the sign dialect most of them use carries few of the technical terms a math or biology lesson runs on. - **The agency moment:** At 19, in a first year of university, Kimathi co-founded ZeroBionic rather than wait for a vendor or a ministry to notice. The hands are designed and built in house out of recycled plastics, printed on 3D machines that run constantly in a small Nairobi lab. Each hand costs roughly $350, and 3D printing brought build time down from two or three months to about a week. The harder gap was not hardware. "We quickly found out that the biggest problem is there's no standardised African sign language dataset," Kimathi told the Daily Nation. "We have American and British sign language, but nothing unified for Africa." So the team stopped trying to write the vocabulary and went to the people who already had it: deaf teachers put on a bodysuit the company imported and recorded the signs themselves, building a database of technical terms that draws on English, American, and Kenyan signs. - **What changed, with evidence:** Machakos Senior Secondary School for the Deaf put forty students on the system this year, evenly split between boys and girls, from Grade 10 to Form Four. Their physics, computer studies, and mathematics teacher, Dennis Ayoyi, told the Daily Nation they "learned about robotics, mechatronics, and even tried basic signs like fingerspelling from A to Z," and that they "were excited and even pointed out both the strengths and weaknesses of the system during the sessions." What Ayoyi rates highest is independence: "It can support online meetings by improving accessibility for learners who rely on sign language communication." What the students caught is the part that still does not work. "The hand shapes were not always clear, and sometimes the robot could not fully close its fist," Ayoyi said. An unclear shape can be read as the wrong sign. Kimathi's own count is more than 78 pilots across Kenya and accuracy at about 92 percent, after an early Rwanda pilot where noise interference made the system pick up and display a vulgar word. - **The move worth copying:** The hand is the hard part to copy. The decision in front of it is not. When the reference they needed did not exist, Kimathi and the team did not commission it or wait for a standards body; they handed a bodysuit to the deaf teachers in the room and let them record the vocabulary, which is why the signs in the database came from the people who sign them daily. If the thing you need written down is already in someone's hands, that is a dataset waiting to be made, and the students who told Ayoyi exactly where the fist failed to close are doing the next round of it. Know someone who built the reference nobody had gotten around to collecting? Feed the scout at goodinprogress.org. - **Takeaway line:** "I started building a small mini phone using Lego, and at the time, I didn't even know what I was doing." — Norah Kimathi, Daily Nation Sources: - Africanews / AFP: https://www.africanews.com/2026/09/09/kenyan-start-up-makes-robotic-limbs-to-sign-for-deaf-kids/ - Daily Nation, Moraa Obiria: https://nation.africa/kenya/news/gender/classrooms-mines-and-maternity-wards-how-three-kenyan-women-are-using-ai-to-save-lives-5424712 Sources last verified: October 2, 2026 --- ## Sylvia Osewe traded evenings of handwritten lesson plans for 30 minutes before class URL: https://goodinprogress.org/stories/sylvia-osewe-kalasik-lesson-plans Location: Kajiado County, Kenya Date: October 10, 2026 Dek: A teacher with up to 200 students a day was close to leaving the profession, then handed the lesson planning to a chatbot and kept the part of the job that mattered. - **The person and the place:** Sylvia Osewe teaches at Oloosirkon Primary School in Kajiado County, in the dry savanna south of Nairobi, and sees up to 200 students in a day. She is also the school's guidance counseling teacher. - **The problem they cared about:** Kenya's public schools were short 98,261 teachers when Rest of World reported the figure in April 2025, and in 2022 about 44 teachers walked away from the profession every day, by the government commission's own count. A shortage like that shows up in a teacher's life as workload. Osewe spent hours of every evening writing lesson plans by hand, on top of the marking and the record-keeping the job already carried. "It was draining, my job became a prison," she told Forbes Africa. Early on, it nearly ended her career. "I wanted to quit; it was very difficult," she told Rest of World. - **The agency moment:** In 2023 a Kenyan nonprofit called Kalasik Education came to Oloosirkon and showed the teachers a chatbot it had built on an OpenAI model and aimed at the Kenyan curriculum. Osewe had assumed none of it was meant for her. "I thought it was for the younger generation," she told Forbes Africa. She started on a phone. And she set her own terms, which is the part that matters, because the tool does not always stay where it is pointed. Kalasik's own chief technology officer told Rest of World the chatbot is supposed to draw only on Kenyan curriculum textbooks and approved sites, and that it sometimes strays onto the wider internet, which the team is working to fix. So Osewe cross-checks the lesson against the textbook before she teaches it, and when a lesson lands badly she asks for a simpler way in. - **What changed, with evidence:** Planning the English, social studies, and Christian religious education lessons, which used to fill her evenings, now happens in about 30 minutes before class. "Since I started using Kalasik, I have more time," Osewe told Rest of World. "The time I used to spend writing lesson plans can now be used to attend to the children." Sixteen months later Forbes Africa found the same arc: hours-long tasks had come down to minutes. "I felt like someone who had won the lottery," Osewe told them. She uses it for quizzes too, and for project rubrics, and, in her counseling role, for behavior contracts and letters home. She keeps a hand on the brake. "AI is a good tool; it makes work easier. We cannot, however, rely on it 100% because it cannot take the place of a human teacher," she told Rest of World. She began on a phone. When Rest of World visited, the school's computer lab held a single desktop from the early aughts under a red Maasai blanket, and the head teacher unlocked a wall cupboard holding six more dated machines, too few, she said, for the school's 1,015 students. Most of Kenya's schools have no internet connection at all. "For those who lack frequent internet access, AI is an idea far from being a doable solution," Bushnell Peter told Rest of World; his job at the Teachers Service Commission is training teachers to use technology. Osewe worries about the access gap closest to her: the students who can reach these tools from home and the ones who cannot. - **The move worth copying:** Osewe cross-checks the lesson against the textbook before she teaches it, and that one habit is the difference between a tool and a shortcut, especially with a tool its own builders say still wanders. If you use one of these things for work that matters, check it against the source you already trust. Share this with the teacher in your life who is still planning lessons at ten at night. - **Takeaway line:** "It makes me feel like I am biased when other students are not able to use the same advantage." — Sylvia Osewe, Rest of World Sources: - Rest of World: https://restofworld.org/2025/ai-teaching-tools-kenya-teacher-shortage/ - Forbes Africa: https://www.forbesafrica.com/cover-story/2026/08/13/learning-that-builds-africas-thinkers-and-innovators Sources last verified: October 10, 2026 --- ## Paula Uzcátegui counted close to 86,000 frailejones from the air, then went back to see which ones lived URL: https://goodinprogress.org/stories/paula-uzcategui-frailejon-census Location: Páramo de Berlín, Colombia Date: October 10, 2026 Dek: A wildfire tore through 317 hectares of Colombian páramo, and counting what was left of its signature plant, one by one, was close to impossible with fieldwork alone, so a student who had studied both biology and systems engineering built the census out of drone images. - **The person and the place:** Paula Uzcátegui was in her eighth semester of systems engineering at the Universidad Industrial de Santander, with two years of biology already behind her, looking for a thesis project that would use both halves of that education. The Páramo de Berlín sits an hour and a half from Bucaramanga, at around 3,600 meters above sea level. - **The problem they cared about:** On January 22, 2024, a fire burned for almost five days and took 317 hectares of the páramo, a figure the regional environmental authority put on the record. Frailejones are the plant the whole ecosystem is built around. When Björn Reu's research group reached the site six days later, the rest of the vegetation was gone and many frailejones were still standing, central leaves intact, and within days some had even begun to flower. Nobody knew how many had come through. The usual way to find out was to count small plots and extrapolate out to the rest, and a full plant-by-plant count across 83 burned hectares was close to impossible with fieldwork alone. - **The agency moment:** Reu's group had the drones and the flight experience. What it did not have was anyone who could make the images give up a number. Uzcátegui took an object-detection model that already existed, cut the drone maps into pieces, and labeled around 12,000 frailejones by hand until the model could find one from above on its own. Hoover Fabián Rueda Chacón, from the university's systems engineering school, is the one who explains what the model learned to key on: the rosette a frailejón makes when you look straight down at it. Cristian Mateo Jaimes Prada flew the surveys, in air thin enough to make the drone labor, with wind, fog, and short battery life squeezing each flight into a narrow window. - **What changed, with evidence:** "En el área grande, de 83 hectáreas, detectamos cerca de 86 000 frailejones con un diámetro mayor a 20 centímetros," Uzcátegui told Mongabay Latam: close to 86,000 plants wider than 20 centimeters, counted individually rather than estimated from a sample. Then, back at the páramo at the end of 2024, the team saw what it had not expected. Frailejones that survived the fire and flowered were dying months afterward. So the team brought in the drone's multispectral sensor and scored every plant the model had found on the balance between the red light it absorbed and the near-infrared it reflected, with a threshold dividing living from dead, and Uzcátegui walked out to plants with a phone and a GPS to check the answer against the ground. The estimate came back at about 30% dead in the year after the fire. "Los frailejones no se murieron por completo por el incendio, murieron tiempo después," Reu said. Andreina Ortiz López, working the same burn from ground level, compared plots inside and outside it and found the number of species already close to matching, with the ground cover still catching up. The 30% is an estimate with a margin of error, and in comparable Andean páramos frailejones go on dying as long as two years out, so one year is a reading and not a final count. The model was trained on a single species, Espeletia standleyana, and Rueda Chacón expects it to carry over, with retraining on a smaller set of images as the worst case. More flights need funding, and the group is hoping to go after it through calls for proposals. - **The move worth copying:** Uzcátegui did not pick between the biology and the engineering, and the census exists because she refused to. The method is cheap enough to run on one flight a year, which is Reu's plan: photograph it again, count again, and watch how a páramo comes back. Know someone whose two unrelated skills turned out to be one skill? Feed the scout at goodinprogress.org. - **Takeaway line:** "Estaba todo completamente negro, chamuscado. Solamente se veían los tallos de los frailejones" ("It was all completely black, scorched. You could only see the stalks of the frailejones.") — Paula Uzcátegui, Mongabay Latam Sources: - Mongabay Latam: https://es.mongabay.com/2026/05/frailejones-drones-inteligencia-artificial-monitoreo-incendio-paramo-berlin-colombia/ Sources last verified: October 10, 2026 --- ## The mountain rescue volunteers who taught a drone what does not belong on a mountain URL: https://goodinprogress.org/stories/binks-roach-parsons-mr-maps-drone Location: The Lake District, England, and Glen Etive, Scotland People: Dan Roach, David Binks, Dan Parsons Date: October 10, 2026 Dek: Charlie Kelly had been missing on a Scottish mountain for six weeks when three Lake District volunteers drove up with two drones and software they had written in their own time. They found him inside the first hour, and his family got an answer before winter. - **The person and the place:** Dan Roach is 34, a climber who flew drones for photography work before ever searching with one, and has been out with Cockermouth Mountain Rescue in the Lake District for about ten years, which somehow still leaves Roach among the younger people on the team. David Binks is 57, volunteers with Duddon and Furness Mountain Rescue, and spent a working life writing simulation code for wind farms and offshore oil rigs. Back in 2007 Binks wrote a program called MR Maps, which let rescue teams watch their own members move across a map in near real time using pings from their radios, and handed it to UK mountain rescue for nothing. Dan Parsons, a friend and teammate of Roach's at Cockermouth, had been running the early experiments with Roach. Roach is the ideas one. Binks is the quiet one with the analytical brain. Mountain rescue in the UK is staffed entirely by volunteers, in teams that are each a separate registered charity raising their own money for their own kit. - **The problem they cared about:** When camera drones got cheap in the early 2010s, everyone in mountain rescue assumed the searching problem was solved. Roach was not convinced. "Everyone thought, ‘This is going to be brilliant because I can see things.’ But then they realized you have to be very close to the thing that you want to see, and you have to be able to see it on a very small controller." The ground is the other half of it. Brian Brathurst, a deputy team leader at Glencoe Mountain Rescue, describes that ground without drama: "There's a lot of boulder fields, there's a lot of gully systems, and you could walk within five or 10 meters of somebody and not see them." - **The agency moment:** Nobody asked for this and nobody funded it. The two of them met over the winter of 2021 into 2022, and Binks came away certain that Roach's idea belonged inside MR Maps rather than alongside it. Then came two years of what Binks calls "some hideous maths." The hours were Binks's own. "I'd start at 9 in the morning, and before I realized, it would be 3 in the morning, and I'd just been at it all day." What Binks ended up with plans a drone's flight path off the Ordnance Survey lidar data that underpins British maps, works out the grid reference of any single pixel in any photo it takes, and then does the part that matters: it counts the pixels of every color in a picture and flags the clusters that are unusual for that picture. Not a trained model, and nothing clever about jackets. Just a count, and an outlier. "Usually it's funny-colored moss, or rocks, or a sheep," Roach says. "But every now and again, it's a person." The whole thing runs offline, because the places it is needed have no signal. Then came the decision that was not technical at all. A search in Scotland had been going for weeks and the three of them had a prototype they had never tested on anything real. "I'd heard this search was going on," Binks says, "and I kept thinking this would be quite a good place to try and test the system out. Then Dan phoned me while I was on my way home from holiday. He'd had the same idea." They called Glencoe Mountain Rescue and asked if they could help. "I don't think they thought anything was going to come of it," Binks says. "So they were happy for us to try." - **What changed, with evidence:** Charlie Kelly was 56, a father of three, and worked as a forensic psychologist inside the Scottish Prisons Service. On September 6, 2023 he left home in Tillicoultry to climb Creise, the peak above Glen Etive. The last message his partner Emer Kennedy had from him put the Glencoe Ski Center lights in view and him back at the car in half an hour. Ten mountain rescue teams searched for him alongside the Coastguard, Police Scotland, and the RAF, with sniffer dogs, quad bikes, helicopters, and conventional drones, at times fifty people on the hill at once. They found his backpack on September 9. After that, nothing, for six weeks. Glencoe scaled back the formal operation after ten days but never actually stopped. "But on our patch, we don't tend to give up on these things," Brathurst says. "We were out every free day we had." On the morning of October 24 the three of them drove up with two drones and split the most promising areas between them, with a Glencoe team member along. Binks started on Area A, the drone drifted out of sight about ten minutes in, and Binks called it back irritated that it had gone wrong, set it flying over Area B, and scrolled through Area A's flagged photos while waiting. "Two of the images had Mr. Kelly in them. I was stunned." It was inside the first hour. Kelly had fallen into a gully, on ground that search teams had walked more than once, and what he had on finished the job of hiding him. "He was wearing a lightish brown top and dark blue trousers," Binks says. "To the human eye that's really hard to pick out." Brathurst was one of the people who had walked it: "We literally walked within probably 20 or 30 meters of him." Binks puts the credit where the six weeks went. "Without the work Glencoe had done, there's no way we'd be able to do what we did." Emer Kennedy got her answer before the winter closed in. "I'm just glad we have closure. It means his kids can have a life." - **The move worth copying:** The software has flown on searches since, mostly to rule areas out, and Parsons keeps the record honest about what that amounts to: "Although we've only found one person, we've never flown over an area where someone was and not found them yet." Binks has never charged anyone for MR Maps and is not about to start, police forces included. "I've never really been interested in monetizing it." But the move worth copying is smaller than the software and costs nothing to borrow. Three volunteers had a half-tested prototype and knew about a search that had run out of ideas, and they phoned a team of strangers before anybody asked them to. The call came before the confidence did. Know someone sitting on a half-finished thing who already knows exactly who it would help? Feed the scout at goodinprogress.org. - **Takeaway line:** "We'd found our coats in fields and simulated it, but we'd never found anything that we didn't know we'd already put there." — Dan Parsons, WIRED Sources: - WIRED, Tristan Kennedy, original reporting and interviews with all three builders: https://www.wired.com/story/this-homemade-ai-drone-software-finds-bodies-when-search-and-rescue-teams-cant/ - BBC News, Steven McKenzie, BBC Scotland News: https://www.bbc.co.uk/news/articles/cw03y9r13pro - grough, Bob Smith, Editor: https://grough.co.uk/magazine/2023/11/04/rescuers-reveal-how-new-drone-technology-helped-find-body-of-missing-glen-etive-walker - Press and Journal, Aberdeen, D.C. Thomson: https://www.pressandjournal.co.uk/fp/news/highlands-islands/6246236/charles-kelly-glen-coe-mountain-rescue/ - Mountain Rescue England and Wales, NON-INDEPENDENT, the subject organization's own statement: https://www.mountain.rescue.org.uk/news/new-lake-district-drone-technology-assists-with-a-difficult-find-in-the-scottish-highlands/ - RICS Land Journal, NON-INDEPENDENT, written by Brendan O'Neill, national drone officer for Mountain Rescue England and Wales: https://ww3.rics.org/uk/en/journals/land-journal/drone-use-mountain-rescue.html Sources last verified: October 10, 2026 --- ## Maria Navarrete addressed the council in Spanish, and the room followed along on two screens URL: https://goodinprogress.org/stories/north-las-vegas-ai-meeting-translation Location: North Las Vegas, Nevada, USA People: Maria Navarrete, Jackie Rogers, Wilson Ramos, Rebecca Gipson, Isaac Barron, Pamela Goynes-Brown Date: October 10, 2026 Dek: North Las Vegas used to send for whichever city employee could translate. In 2023 it put a laptop on the speakers' podium instead, and by 2026 the city clerk says there is almost no language you cannot bring to City Hall. - **The person and the place:** In North Las Vegas, the most recent census had 42 percent of residents identifying as Hispanic and almost 38 percent speaking a language other than English at home. Maria Navarrete lives there and belongs to Make The Road Nevada. - **The problem North Las Vegas cared about:** Public comment is the plainest tool a resident has, and it only works if the room can understand you. The city used to handle that by sending for whichever city employee could translate when the need arose. Rebecca Gipson, the assistant city manager, said the city wanted something more efficient that could also serve the residents watching from home. Gipson put the ceiling on the old arrangement: "We'd have to have 30 interpreters in the council chambers … it would be a little bit impossible to do that in an in-person format." - **The agency moment:** The city bought Wordly, and in August 2023 it debuted real-time AI translation at its public meetings, the first jurisdiction in Nevada to do it, and set it up for the room rather than for a press photo. A laptop went on the speakers' podium so that anyone at the microphone could watch the translation while talking. Council members got it on their own screens. Two televisions in the chamber carried the transcript, one in English and one in Spanish. Anyone in the audience, or at home with the agenda open, could scan a QR code and read the translation on a phone, or listen to it on earphones instead of reading it. It started with Spanish and English. - **What changed, with evidence:** At one of the first meetings, Navarrete walked to the microphone and thanked the council in Spanish while Wordly put the words on the screens in English. Councilman Isaac Barron told the room: "As a teacher and member of the community, I'm overjoyed." Wilson Ramos, who runs community services and engagement for the city, named the part that actually moves: "Many of our community members have never been to a council meeting before." Mayor Pamela Goynes-Brown, in the same meeting, gave the honest version of how new it felt: "What I like about this is it's translating English to Spanish and Spanish to English and it's not missing a beat at all. We just have to get used to reading the screen over here." The translation is not flawless either. Accuracy depends on how fast and how clearly a person speaks, and the audience can watch the software reach back and clean up a sentence mid-thought. It was still running in March 2026, by which point Government Technology reported Spanish as the default with other languages reachable by QR code, and City Clerk Jackie Rogers put the range this way: "There's almost no language that you can't come to City Hall and speak." Tagalog, which officials said in 2023 they expected by the end of that year for the city's growing Filipino population, was pending a software update from the company. - **The move worth copying:** The equipment was a laptop, a QR code, and the two televisions in the council chamber. Dozens of cities have put similar systems in place, particularly in California and Nevada. If you know a city clerk, a school board member, or anyone who runs a public meeting where part of the room is not following what is said, this is the one to put in front of them. - **Takeaway line:** "La tecnología de traducción es algo más que una comodidad, es una herramienta vital para gente como yo. (Translation technology is more than just a convenience, it's a vital tool for people like me.)" — Maria Navarrete, to the North Las Vegas City Council, in The Nevada Independent Sources: - The Nevada Independent, Jannelle Calderon, 2023: https://thenevadaindependent.com/article/north-las-vegas-is-using-ai-to-translate-public-meetings-in-real-time - KTNV Channel 13 Las Vegas, 2023: https://www.ktnv.com/news/north-las-vegas-using-ai-to-break-language-barriers-in-public-meetings - Government Technology, Carl Smith: https://www.govtech.com/artificial-intelligence/ai-lets-state-and-local-governments-extend-their-reach - Government Technology, Jule Pattison-Gordon: https://www.govtech.com/civic/cities-bring-live-automated-translation-to-public-meetings - FOX5 Las Vegas / KVVU: https://www.fox5vegas.com/2023/07/31/city-north-las-vegas-launch-ai-powered-translator-city-council-meeting/ Sources last verified: October 10, 2026 --- ## Sateesh Nori wanted a chatbot for tenants. The team picked the one topic where being wrong wouldn't get anyone evicted. URL: https://goodinprogress.org/stories/sateesh-nori-roxanne-repair-bot Location: New York City, USA People: Sateesh Nori, Jenny Laurie, Sam Flynn Date: October 10, 2026 Dek: New York's right to a lawyer stops at eviction, and the nonprofit hotline that picks up the rest cannot reach everyone who needs it. So a team of about five volunteers spent 13 months teaching a bot one corner of tenant law. - **The person and the place:** Sateesh Nori spent years doing legal services work in New York City and now teaches at NYU School of Law. The arithmetic was familiar from the inside. "One of the biggest challenges of working as a legal services lawyer or a nonprofit lawyer is that you don't have enough time, or resources, to help everybody who seeks your help," Nori said. - **The problem Sateesh cared about:** New York City gives tenants a right to a lawyer when they are facing eviction. That right stops there. A tenant with no heat, no hot water, a dead elevator, or mold on the wall calls Housing Court Answers. By Nori's account it is a nonprofit with a staff of 15 and a budget of about a million dollars, its hotline takes about 200 calls a day, and the wait to reach a person runs 45 minutes to an hour. The other option Nori describes is a trip to a courthouse or a legal services office, sometimes three to four hours on a bench, and then usually being turned away for not facing eviction. - **The agency moment:** Nori had wanted to point AI at tenants' problems since 2023, and the first decision was which problem. Eviction and rent were the obvious candidates, and the team ruled them out on purpose. "We thought about eviction and rent issues, but we decided against that, because if a chatbot makes a mistake, we didn't want to risk somebody getting evicted," Nori told Law360. They took repairs instead, and trained the bot to answer "I don't know" the moment a question left that subject. - **What changed, with evidence:** Nori matched Josef Legal with Housing Court Answers, making three organizations with NYU Law School, and a team of about five people spent 13 months feeding it every statute, guide, and training document they could find about repairs in New York City. Nobody funded it and everybody had another job. Sam Flynn, who co-founded Josef, says Housing Court Answers, Josef, and Nori spent about six months testing it and checking its answers. Roxanne launched in January 2025, usable from a phone, drawing on Housing Court Answers' own guidance instead of the open internet. Brick Underground asked it how to get a landlord to fix the heat and got the real sequence back: make a record, tell the landlord, call 311, talk to the other tenants, and file an HP Action in housing court if the repairs still are not made. The same test found it will occasionally answer something outside its subject. Jenny Laurie, who runs Housing Court Answers, states the rest plainly: "This is a machine. It's not a live lawyer, so you're not getting legal advice that's custom made for your situation." Over three to four weeks Nori counted at least 1,000 uses. A lot of them were about heat. It was a cold January. - **The move worth copying:** The team did not start with the biggest problem. They started with the one where a wrong answer was survivable, and then trained the thing to refuse everything else. If you are pointing AI at something that matters, that is the first call to make: find the corner where being wrong costs the least, and begin there. - **Takeaway line:** "For us, it's not even about the AI, it's about the results." — Sateesh Nori, to Law360 Sources: - Brick Underground, Celia Young: https://www.brickunderground.com/live/housing-court-answers-josef-chatbot-roxanne-answers-nyc-tenant-repair-questions - Law360, Marco Poggio: https://www.law360.com/articles/2293691/nyu-law-professor-on-his-ai-powered-tenants-rights-bot - LawNext, Bob Ambrogi, January 2025: https://www.lawnext.com/2025/01/ai-powered-tool-launches-to-help-new-york-tenants-enforce-their-repair-rights.html - Thomson Reuters Institute, Natalie Runyon: https://www.thomsonreuters.com/en-us/posts/ai-in-courts/housing-repairs-chatbot/ - Housing Court Answers, the tool's own page: https://housingcourtanswers.org/roxanne/ Sources last verified: October 10, 2026 --- ## The alpine commander who would rather a drone guess out loud than say nothing URL: https://goodinprogress.org/stories/phillip-eberle-frnsw-ai-drone-rescue Location: Kosciuszko National Park, New South Wales, Australia People: Phillip Eberle, Russell Turner, John Marzol Date: October 11, 2026 Dek: Police called for help from the air on a freezing alpine night, and the Jindabyne crew that answered flew a drone that speaks up while it is still unsure, with two hikers out of the park in under five hours. - **The person and the place:** Phillip Eberle is an Inspector with Fire and Rescue NSW and the regional duty commander for the New South Wales alpine area, which makes the Snowy Mountains in June Eberle's patch. The fire station that matters here is at Jindabyne, the last real town before the high country, about 35 kilometers from the Dead Horse Gap walking track, which lies to the southwest. The drone belongs to the FRNSW Aviation Unit, whose team leader and chief remote pilot is another Inspector, Russell Turner. Turner puts the aircraft inside a statewide upgrade that has been arriving for about 18 months. Each one carries four cameras, infrared among them, and software that reads thousands of images in real time, and the same aircraft can lay roads, bush tracks, and fire trails over the operator's screen while a search is running. The people who flew one that night were not specialists sent up from a city. They were the local firefighters at Jindabyne, and in FRNSW's own account of the rescue, Acting Superintendent John Marzol credited exactly that: "The training and support provided by FRNSW's Aviation Unit ensured local firefighters had the skills and equipment to support Police and the SES to effect a successful operation." - **The problem they cared about:** On the evening of Tuesday, June 23, 2026, two hikers in their early 20s walking the Dead Horse Gap track failed to reach their rendezvous point. They were reported missing at 7pm. By then it was dark and cold, and they had drifted some 500 meters clear of the marked route, which in broken alpine country is far enough to be invisible. NSW Police put an Alpine Search and Rescue team on the ground, with State Emergency Service volunteers alongside them, and called Fire and Rescue NSW for help from the air. The arithmetic of a night search in the snowfields is not only about the people who are lost. Every hour it runs is another hour of volunteers out on a freezing mountain. - **The agency moment:** The call went to the Jindabyne station, and from the command post the crew put a remotely piloted aircraft up and started flying a grid. What the software on board actually does is narrower than it sounds, and it turns on a decision about doubt. "The AI scans the area and scans the information that it can see on its camera and says, 'OK, this item that I'm looking at has a 35 per cent chance of being human'," Eberle told About Regional. The instinct is to want a higher bar than that. Eberle's point is that the higher bar is the thing that loses people. "If we'd mark that as being, say, 100 per cent [sure that it was a human], it wouldn't pick it up," Eberle said. "It would say, 'No, I can't 100 per cent say that's a human, so I'm not going to tell you about it'." So, on Eberle's telling, it speaks while it is still unsure, and the speaking is not the end of it. "It'll tell the pilot, 'This is a person', and then we'll fly over and double-check that it is," Eberle said. Turner told ABC News what that looks like from the pilot's seat: "When the machine determines that it thinks it's a person or a vehicle or a vessel, it will ping and put a display on exactly where it thinks that person is." Nobody on that mountain was taking the machine's word for anything. They were taking its guesses seriously enough to go and look. - **What changed, with evidence:** Working the grid from the command post, the crew got their flag: the software had picked the pair out of the bush, roughly half a kilometer from the track, and the thermal camera zoomed in to confirm it. From the ground, the hikers were signaling too, holding up a mobile phone with a red light on it to pull the drone toward them in the dark. The operator answered with the drone's spotlight to point the ground crew at the spot, and police and SES volunteers used the aircraft's built-in speaker to talk to the pair before walking them off the mountain. Eberle's account has the ground teams taking about four hours to reach the pair after the find, and they were walked to safety around midnight. Both came off the mountain with mild exposure, and neither accepted treatment there. FRNSW called it a first-of-its-kind mission, and it was the first time the service's drone AI detection system had located a missing hiker. Eberle counts what it saved in days rather than hours: "It's definitely helped make what could have been a long-term incident into a very short-term incident." And on how else the night could have gone: "There's every chance in the world we could have still been out there." - **The move worth copying:** Almost nobody reading this is going to put a drone over a mountain at night, and the transferable part is smaller than the aircraft anyway. It is a habit about hedging. The instinct with any tool that reports a probability is to tune it up until it only speaks when it is certain, and on Eberle's account that is precisely the setting that would have reported nothing. The practice Eberle describes runs the other way: let the software flag what it half-sees, and send somebody to confirm it. The night belonged to a lot of people. The police who called for air support, the SES volunteers who walked in, the Jindabyne crew who flew, and two hikers who had the sense to hold up a light. Know a crew, a clinic, or a service that treats a machine's uncertainty as information instead of a fault? Feed the scout at goodinprogress.org. - **Takeaway line:** "It means less time on the mountain and less time looking for people, and reduced risk so that the volunteers can get home to their families quicker" — Phillip Eberle, ABC News Sources: - ABC News, Mikayla McGuirk-Scolaro, ABC South East NSW, original reporting with separate interviews of Inspectors Eberle and Turner: https://www.abc.net.au/news/2026-06-27/lost-hikers-found-by-ai-drone-in-kosciuszko-national-park/106846712 - About Regional: https://aboutregional.com.au/how-ai-saved-two-hikers-lost-in-the-unforgiving-alpine-wilderness/509715/ - Fire and Rescue NSW, NON-INDEPENDENT, the responding agency's own media release, 2026-06-25: https://www.fire.nsw.gov.au/media/news/2026/20260625-fire-and-rescue-nsw-drone-team-locates-missing-hikers-using-ai-technology Sources last verified: October 11, 2026