← Projects

LedgerLens

Document AI

Bank statements and scanned cheques come out as a reconciled ledger, and nothing counts as done until deterministic validators pass: MICR checksums, written-versus-numeric amounts and the closing balance.

Screenshot of LedgerLens: Interface illustration with results from the repo’s own validators: MICR, written amounts and closing balance pass; handwriting goes to review.
Interface illustration with results from the repo’s own validators: MICR, written amounts and closing balance pass; handwriting goes to review.

Highlights

The full story

Bank statements and scanned cheques to a reconciled ledger. Printed fields come straight from the PDF’s text layer, handwriting goes to OCR plus a vision-language model, and deterministic validators check MICR checksums, written-vs-numeric amounts and the closing balance before anything counts as done. Per-bank adapters, an editable grid with Excel export, and an invoice VAT/GL coder share one multi-tenant platform. The public repo is a clean-room version on synthetic data.

Built with

Python · FastAPI · Mistral OCR · Pixtral · PyMuPDF · Supabase