The IMDb of AI — built by AI.
A live, citation-first encyclopedia of the AI industry — 501 companies, 272 people, 21 categories — refreshed continuously by a swarm of Claude research agents that show their sources.
The problem
The AI industry moves faster than any analyst desk can. Crunchbase profiles are six months stale. Pitchbook is paywalled and still slow. Wikipedia is volunteer-paced. And every article a journalist writes about an AI company is already wrong by the time it publishes — because the company shipped a new model, raised another round, or hired three new VPs that week.
There was no neutral, citation-first reference that an operator, investor, or journalist could check first. The opportunity was to build it — but the only way the math worked was if the agents did the research, not the analysts.
What we built
AIDB is a Postgres-backed encyclopedia of AI companies, models, and people, where every claim — valuation, ARR, headcount, funding round, founder transition — links back to a primary source with a retrieval date. The freshness of every profile is shown as a colored badge: green under 30 days, amber under 90, red older. Compare two-to-four companies side-by-side and the deltas surface visually.
Behind the front-end, a fleet of Claude Opus 4.6 research agents runs continuously. Each agent owns a topic — a company, a person, a category — and is responsible for keeping its corner of the database fresh. The agent fires Anthropic web searches against a curated source list, parses returns into structured fields, checks them against the existing record, and writes a diff with attached citations into a versioned log. A human reviewer approves anything material; everything else lands directly.
The daily five-brief newsroom on the front page is the same swarm reading its own log and selecting the five most material moves of the last twenty-four hours.
The first reference I check on any AI company now is AIDB. Wikipedia is months behind, Crunchbase needs decoding. AIDB shows the source and the date next to every number.
How it shipped
Two weeks of schema work — landing on a normalized model that could store time-series facts ("ARR was $10M as of March 2026, cited from the company's Series B announcement") rather than overwriting values. Three weeks of agent prompt engineering, calibrating against a manually-curated gold set of fifty companies until the agents agreed with the analyst on every material field. Two weeks of UI — the compare view, the freshness badges, the citation popovers, the brief. Two weeks of supervised launch where every agent write was human-approved so the prompt-and-source library could be tuned without anything bad getting into the database.
Pricing is a freemium read with paid API access for institutional users. The marginal cost of a profile refresh is the inference plus the search — fractions of a cent — so the read economics scale linearly.