Document Intelligence Suite
Three parsers reading invoices, emails and timesheets past 90% field accuracy, including the ones that arrive as photos of paper.
Three production tools sharing one extraction backbone:
- Scanned invoice parser. OCR plus layout-aware extraction. Half the input is a phone photo of a creased page, taken at an angle.
- Email analyzer. Thread-aware classification and field extraction. The fact you need is usually three replies up, not in the message you were sent.
- Timesheet parser. Structured output from anything between a clean table and a paragraph of notes.
All three clear 90% field-level accuracy. They run as async FastAPI endpoints, which was deliberate: other teams integrate against a stable HTTP contract and never have to know which model is behind it. Docker deployments, CI/CD on AWS.
The accuracy number is what gets quoted, but the thing that actually moved it was cutting scope. Each parser got much better once we stopped trying to pull every field off the page and only extracted the ones downstream systems actually read.
innoscripta SE · Led development