Twenty years of scanned documents, made queryable
Turning an archive nobody could search into structured, checkable records.
What we found
[00]% of documents were parsed to structured records with no human review; the remainder were routed to a review queue rather than guessed at.
Most businesses of a certain age have an archive like this: filing cabinets that became folders of scans, holding information the business genuinely needs and cannot query.
The interesting problem is not extraction, it is confidence. A pipeline that returns an answer for every document will quietly return wrong answers for the hard ones, and nobody notices until a number is off.
We built the pipeline to know what it does not know: pages below a confidence threshold go to a review queue with the crop that caused the doubt, so a person spends their time only where it is needed.
The output is an ordinary database the client's team can query with ordinary tools. That is the point — the value is the structured records, not the model that produced them.
Draft — re-run and replace [bracketed] figures before publishing (lib/lab.ts)