Document Intelligence Suite
Three parsers reading invoices, emails, and timesheets past 90% field accuracy — including the ones that arrive as photographs of paper.
Three production tools on one extraction backbone:
- Scanned invoice parser — OCR plus layout-aware extraction, on documents that are frequently a phone photo of a creased page.
- Email analyzer — thread-aware classification and field extraction, where the relevant fact is often three replies up.
- Timesheet parser — structured recovery from formats ranging from clean tables to freeform notes.
All three clear 90% field-level accuracy. They’re served as async FastAPI endpoints, which was a deliberate choice: other teams could integrate against a stable HTTP contract without knowing or caring which model was behind it. Docker deployments and CI/CD on AWS.
The accuracy number is the headline, but the useful lesson was about scope — each parser got dramatically better once we stopped trying to extract every field and focused on the ones downstream systems actually consumed.
innoscripta SE · Led development