Invoice Reader

Problem

Bookkeeping still means retyping invoices. OCR alone misreads totals and can’t tell line items from footers.

Try it live

Pick a sample, such as a clean PDF, a crumpled photo, a handwritten receipt or a euro invoice with VAT, or upload your own. Each field lights up on the document with its confidence. Fix a field, then export CSV or JSON.

Proves

Vision models with typed output, and validation that keeps arithmetic away from the model.

Signal chain

1. OCR words + boxes
2. screen for injection
3. extract to schema
4. validate and reconcile
5. repair once
6. map to boxes
7. check duplicates
8. journal entry
9. export

Stack

Flask + an asyncio pipeline · OpenAI-compatible client to the gateway · Pydantic v2 structured output · RapidOCR (ONNX) in a locked-down process · Cloudflare R2 · Vue viewer

Engineering highlights

• The pipeline is async because every step waits on I/O: the model, storage and the OCR worker. Two documents in flight overlap, and a test proves it under the production server.
• Deterministic checks own the math. Line items must sum to the subtotal and VAT must match its rate, or the document goes back for one targeted repair. A document that still fails is returned with its failing check, never fixed silently.
• Duplicate invoices are caught by vendor, number and content hash before export.