Invoice OCR: Automatic Invoice Reading Into Structured JSON

A language model can return an amount that does not appear anywhere on the invoice. That sounds harmless until the amount lands in your ledger. So for every value we read, we attach the exact fragment of the document it came from and the coordinates of that spot on the page. If a number has no backing in the scan, we do not pretend it is fine. It gets a confidence of zero and goes back to a person together with the image crop, so checking it takes seconds instead of a search through the PDF.

The second check is plain arithmetic. Line items have to add up to the net amount, VAT rates to the tax amount, and net plus tax to the gross total. An invoice that fails this test does not slip through quietly. It comes back with a note on what does not add up and where. In practice this is the test that catches most mistakes before anyone sees them in the accounting system.

General models read invoices well. A model tuned on one company's documents reads them better, because it knows that company's suppliers, page layouts and the way line items are described. In 2026 studies a small open model tuned this way outperformed the largest commercial models, and the gap was widest in the field that causes the most trouble: dates.

We run the training on our own GPUs in Poland, on a few thousand of your invoices, without sending them to any external API. The finished model works for you alone. We quote the timeline and cost after looking at a sample of your documents, because both depend on the number of templates and the quality of the scans.

Accuracy on clean PDFs says little about what happens to a photo of an invoice taken in a warehouse in poor light. So before rollout we take a sample of your documents and deliberately degrade it: blur, skew, creases, compression from a messaging app.

You get a report with accuracy for every field and every kind of damage, along with the threshold below which a document should go to manual review. We rerun the same set after every model update, so you know the quality has not dropped quietly.