You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
ATLAS is a map of everyone in AI safety. Each person is a face on a real world map, coloured by their research field and badged by what they want right now: hiring, looking for a job, looking for collaborators, or looking for funding. Press "Find for me" and it matches you both ways. A job-seeker watches arcs reach out to the labs, mentors and programs that want them. A lab presses it and the matching people get pulled in from across the world into a ranked shortlist.
The field fills itself. A scraper pulls people from LinkedIn, personal sites, GitHub, arXiv and the Alignment Forum and drops them on the map straight away, before they ever sign up. AI writes a first draft of each profile from their public work and works out their field and skills. Then anyone can log in, claim their pin and edit it.
A full demo is already live at aisa.nahdha.tech: six surfaces, 1,020 people, a working match engine, deployed.
The idea comes straight out of Austin Chen's "Sixteen schemes for AI safety" -- scheme #1 (Triplebyte for AI safety jobs) and #2 (a database of every AI safety person). His point is the whole reason to build it: money is pouring into AI safety, so money stops being the bottleneck and finding the right people becomes the bottleneck.
The field is impossible to see. Recruiters can't find talent, newcomers can't find roles or mentors, people can't find collaborators, independent researchers can't find funders, and nobody can see the shape of the field where the work is and where the gaps are. People have tried to fix this with databases and org-maps like aisafety.world. Databases rot because nobody updates a spreadsheet, and those maps show orgs, not people, and they don't match anyone to anything.
The goal: kill the search cost for all five of those at once, with real people, and put the global south on the map instead of treating it as an afterthought.
How it works, and why it won't rot like the rest:
The scraper. The map is populated automatically from public sources, so you don't wait for people to show up they're already there. This is Austin's #2, almost word for word: scrape LinkedIn, socials and personal sites, then let people edit.
AI does the boring part. Each scraped person gets a skill vector computed from their public work, a draft bio, and suggested fields and tags, so claiming a profile is basically one click. Search is plain English: "someone working on chain-of-thought faithfulness, open to collaborate, based in Africa." And because the field is computed from real work, the empty spots on the map show where the field isn't working yet.
It's a matchmaker, not a list. One cosine-similarity engine over a ~99-skill vocabulary runs Find-for-me, the skill galaxy, the project beacons and the speaker finder.
Three-month plan:
Month 1: Build the real backend with logins. Build the scraper and get real people onto the map automatically. Opt-out and privacy controls in place.
Month 2: The AI layer: embed each person's public work to compute their field and skills and draft their profile; the claim-your-profile email and login flow; plain-English search; Find-for-me running both directions on the real population.
Month 3: Scrape orgs too (Austin's related idea, a database of every AI safety org), wire the live, community and conference surfaces and the speaker-finder to real data, polish, and launch.
Three months of me building it full time, plus what it costs to scrape and run it on real data.
My time: $1,500/month × 3 = $4,500
AI and scraping embeddings, LLM for profiles and search, scraping compute about $1,200
Servers, database and the email service for claim-your-profile invites it's about $1,200
Domain is about $100
Just me. Volunteers welcome if anyone is interested.
The best evidence I can build this is that I already built it. I designed and shipped the whole ATLAS demo alone: six surfaces (the map with live clustering of 1,020 people, a skill-similarity galaxy, a directory, live rooms, community rooms, conferences), a working match engine, and a full design system, deployed and running at aisa.nahdha.tech.
Background: around 4 years of applied ML and NLP engineering. MSc in Mathematical Sciences (AI for Science) at AIMS South Africa / University of Cape Town on a Google DeepMind Scholarship, thesis on neural reasoning for ARC-AGI. Research engineer on ARC-AGI at Peking University, Nov 2025 to Feb 2026. At Sultan Qaboos University I built a proposal-evaluation NLP system using multilingual embeddings scored against Oman Vision 2040 the same embedding and retrieval work the matching engine runs on. Hackathon wins: first in Qatar, first in Kigali, third at the Deep Learning Indaba. My interp work is public at github.com/AhMedDa1/mech-interp-journey.
And I'm the user. I work from Sudan, Rwanda, Oman, outside every AI safety hub, with no ready-made network. Finding mentors, collaborators and funding from here is exactly the problem this fixes. I'm building the thing I needed and couldn't find, for the people in the same spot.
Most likely ways it fails:
Nobody claims their profile. The scraper means the map is never empty, but if people don't come back to claim and edit, the data drifts. Fix: AI-drafted profiles make claiming nearly free, and it's built to be worth looking at, so people want to be on it.
Consent and privacy. Putting people on a public map is sensitive, and some people in this field keep a low profile on purpose. Opt-out and removal work from day one, profiles default to public-information-only, and anyone can take themselves off.
If it fails anyway the downside is small. The scraped public map stays up as a free reference, the code and match engine are reusable, and I'll know what makes people claim a profile and what matches they actually want, which is useful to whoever tries next.
$0 for this project. I paid for the demo and hosting myself.