DocuMind
Document AI for invoices, receipts and bills. Snap a photo, a scan or a PDF and DocuMind reads the vendor, NTN, document number, dates, line items, sales tax and total, checks that the numbers add up, and flags duplicates and tampered totals.
- Page detection, flattening and deskew with OpenCV, then neural OCR (docTR) on the GPU
- A layout transformer trained from scratch labels every word on the page: 96.8% of fields exactly right on 600 new Pakistani-style documents
- 77.2% of company, address, date and total exactly right on real receipts from the SROIE test set it never trained on
- Cross-checks every document: line items add up, subtotal plus GST equals the total, the GST rate is real, NTN format and dates
- Recognises well-known issuers such as PTCL, K-Electric, LESCO and SNGPL
- Dashboard with spend insights, duplicate and unusual-amount flags, and export to JSON or Excel
- Batch mode reads many documents at once
Built with
- Document AI
- PyTorch, docTR OCR, Tesseract, OpenCV
- Language AI
- Self-hosted LLM
- Backend
- Python, Flask, SQLite, PyMuPDF
- Frontend
- React, Vite, Framer Motion