Banking / compliance · Jun-Jul 2026
On-premise document verification & field extraction
On-premise service for a Tunisian bank's securities-order and fund-subscription forms. Given a scanned PDF, it verifies the four mandatory marks - agency stamp, bank signature, "signature conforme," client signature - extracts ~30 structured fields, and returns a three-tier verdict with an explainable 0-100 trust score.
What it did, and what it measured
- 96.1% extraction accuracy on the demo set and 91.3% on a held-out 115-document test split, reached across six documented hardening rounds, each gated by a zero-regression check against the previous scorecard.
- Designed a hybrid extraction strategy - regex-first for the 16 fields with proven coverage, zone-OCR-first for the 14 fields regex never handled. Lifted
nationalitefrom 45.2% to 81.7% anddate_bulletinfrom 0% to 58.7% without regressing a single existing field. - Built the full evaluation harness (detector boxes, extraction, rejection, trust calibration) producing committed scorecards, plus a synthetic document generator - so every published number traces to a file.
Stack
Python 3.12PyTorch 2.5 + torchvisionOpenCVTesseractpypdfium2FastAPIPydantic v2Streamlitpytest
Repo and demo
This is client work, so the source is not public. Happy to walk through the architecture, the scorecards and the trade-offs on a call.