Engineering study 03
Extraction that returns null
A field the model is not sure of comes back empty and flagged, never filled with something plausible.
KycCheckAITypeScriptPublic repository
The problem
Document extraction fails in the most expensive way possible: it returns something that looks right. A plausible value in a typed field passes validation, passes review, and enters your system as fact.
INIdentity documentimage or PDF
EXTRACTTyped extractionschema-bound
CONFIDENCESure or notnull if not
REVIEWReviewer consolea person decides
Null is a valid answer. Prompts are written to return null rather than a plausible value when the source does not clearly say it. The uncertainty is surfaced to a reviewer instead of being hidden inside a confident-looking record.
Hybrid matching, because neither method works alone
Matching an extracted company name against a register needs two mechanisms at once: vector similarity for meaning, and fuzzy string matching for near-identical spellings.
Neither alone resolves “Northgate Holdings LLC” against
“Northgate Hldgs, L.L.C.” Vector similarity understands they mean the same thing
but scores dozens of unrelated companies nearly as highly. Fuzzy matching catches the abbreviation
but misses a genuine rename. Run together, they resolve it.
Built on synthetic fixtures. No real identity data has been
processed.
Where to go next
Bring us one difficult document
Sixty minutes on your extraction workflow, and an honest answer on where null beats a guess.
A 60-minute architecture review · no charge · the notes are yours either way