Document processing

Output is produced against a schema rather than parsed back out of prose. A field that must be a date is typed as a date, an amount is typed as an amount, and a response that does not conform fails validation at the boundary instead of moving downstream to be discovered by whatever breaks first.

Every extracted field carries a confidence, and a threshold routes the record: above it, the value is accepted automatically; below it, the record enters a review queue. Where that threshold sits is a business decision about the cost of a wrong value weighed against the cost of a person looking, and it is set deliberately, not left at whatever the library ships with.

The review queue is designed as a real interface, not treated as the part that does not need design. The document sits beside the extraction, the uncertain field is highlighted, and accepting a value takes a keystroke. Where reviewing a record takes as long as retyping it did, the system has moved the work rather than removed it, and it will be abandoned within a quarter.

Corrections made in review are captured as data. A person fixing the same field the same way every day is a signal about either the extraction or the schema, and a correction that disappears into a database without being counted teaches nobody anything.

Documents that should not be processed at all are detected and rejected rather than processed badly: the wrong form entirely, the illegible scan, the photograph of a screen, the page that arrived rotated. Rejection with a reason is a supported outcome, and a pipeline whose only options are a result or a crash will produce confident results from unreadable inputs.

Volume and unit cost are measured before anything is committed to. Per-document processing cost against per-document manual cost, at the volume the business actually sees, not the volume in the proposal, is an arithmetic question with an answer, and it occasionally returns the answer that the manual process is cheaper.

Personal data in documents constrains where processing may happen. What a vendor retains, for how long, whether the content is used for training, and whether it may leave a jurisdiction are contract questions that precede the technical ones, and they are settled in writing before a single document is sent.

Quality is reported per field rather than per document. A printed total and a handwritten margin note are not the same task and do not fail at the same rate, and one aggregate figure across both conceals which of them is actually costing money.