AIDB

Executive summary

AIDB is Digital Boutique's answer to a structural problem in AI research: the industry moves faster than any hand-maintained reference can track. Valuations change in a week. Companies pivot in a month. By the time a directory is compiled it already misrepresents its subject.

The build treats currency as an architectural property rather than an editorial chore. Every record carries its sources and the date each was retrieved, and an autonomous research agent can refresh any entity against live web search on demand — returning strict-schema JSON, a citation trail and a changelog. The result is a public index of 960 companies and 501 people that reads like a paper and refreshes like a feed.

The challenge

Two kinds of product already existed and both failed the same test. Directories are comprehensive but stale: they present a snapshot with no indication of when it was taken, so a reader cannot tell a fresh figure from a year-old one. News feeds are current but unstructured: they tell you what happened without maintaining the entity underneath, so there is nothing to sort, rank or compare.

Neither is usable as a source of record. Being right once is not the same as staying right, and a reference work that cannot show its age is worse than no reference at all — it launders stale data as current fact.

The requirement was a system that is encyclopedic and current at the same time, and that proves which of its claims can be trusted. That meant three commitments that shaped everything else: every figure traceable to a primary source with a retrieval date; every profile displaying its own verification age honestly, including when that age is bad; and refresh cheap enough to run continuously rather than in periodic manual sweeps.

Approach

The refresh problem is what makes the rest tractable, so it was solved first. A Claude Sonnet 4.6 agent takes a single entity, runs live web search, and is constrained to emit strict-schema JSON — sources, claims and changed fields, nothing else. Because output is schema-bound rather than prose, results write straight into Postgres and every run produces a diffable changelog. The same agent runs from a button on a profile and from a cron across the index; there is one code path, not two.

Freshness is then a rendering concern rather than a maintenance one. Each profile stores when it was last verified and surfaces that age in the interface — green under 30 days, amber under 90, red beyond. The interface is candid about its own staleness, which is what makes the green states worth believing.

Citations are treated as a contract, not a footnote. Each figure links to the primary source it came from with the date it was retrieved, so a reader can check any claim without leaving the profile.