Docket: AI Invoice Intake with a Fraud Gate
Krystian Siek
Docket turns incoming invoices and documents into checked, structured data, for Europe and the US.
What it does
Every extracted value is traced to its place on the page.
Scans and phone photos are read twice, by a vision model and by OCR, and the results are compared.
Arithmetic, VAT, US sales tax, IBANs and US routing numbers are checked in code.
A risk gate stops bank-detail swaps, look-alike suppliers, duplicates and PO mismatches before anything is paid. The model reads, code decides.
Published eval
99.2% of fields right and 98% of documents routed right, across 8 countries, PDFs and phone photos.
About 3 s median per document.
Stack: Python, TypeScript, PostgreSQL, Docker, OCR + vision model.
Try it with your own invoice: https://adaptivesystems.dev/docket/try
Like this project
Posted Nov 5, 2026
Invoices, scans and phone photos into checked data for Europe and the US. Every value traced to the page, totals and bank details checked in code. Published eval: 99.2% of fields right, about 3 s per document.
Likes
0
Views
3
