Upload an invoice and watch an AI model pull out the fields, quote where it found each one, and let plain code check that the numbers add up.
drop an invoice
PDF, PNG or JPG
up to 4 MB · up to 5 pages
Northwind
northwind-invoice.pdf
3 KB · sample
Before the AI sees it
Really a PDFchecked by its first bytes, not the name
3 KB, under 4 MB
1 page, up to 5 allowed
Within the hourly limit
in memory onlyNo copy is stored.
01Upload an invoiceOne PDF or image, checked first
Sample data · illustration
Transcript of the animation.
Upload an invoice (One PDF or image, checked first). A sample invoice from Northwind Tools is dropped on the reader. Before any AI is involved, the server checks the file by its contents: it really is a PDF, it is under 4 MB and it has one page. Nothing is stored.
Read text and layout (Digital PDF or scan). A digital PDF carries real text with positions, so it reads exactly. A scan is only a picture of a page, so the model has to read the pixels, and a skewed page or a stamp makes mistakes more likely. Either way the model sees the layout: what sits next to Total, which column is the amount.
Fields, with evidence (Every value quotes the page). The model fills a fixed set of fields. Each value comes with a confidence level and a short quote from the document as evidence, highlighted here on the page. A field that is not printed stays empty instead of being guessed.
Plain code checks the maths (No AI in this step). After extraction, ordinary code checks the result: line amounts, lines against the subtotal, subtotal plus tax against the total, dates and currency. On the Fernway scan, 520.00 plus 93.60 is 613.60, but the total was read as 618.60, so the check fails and the document is not passed on.
Unsure? Ask a person (Low confidence is flagged). A date printed as 03/04/2026 could be the third of April or the fourth of March, so the model marks it low confidence. The reviewer sees the quote beside the value, picks the right reading and confirms.
Clean JSON out (Ready for your accounting system). The checked result comes out as clean JSON: every field with its value, confidence and evidence, the line items and the check results. It can be downloaded, or in a production system sent straight to the accounting software.
An illustration with fictional sample invoices. The live tool is right below.
Try it
Read an invoice now
One PDF, PNG or JPG at a time. Or pick one of three fictional sample invoices.
Files are processed by a third-party AI model (Google Gemini) and are not stored by us. Do not upload confidential documents; use the samples if unsure. Need this on your own servers? See private AI.
The server confirms it is really a PDF, PNG or JPG by its first bytes, under 4 MB and at most 5 pages. The file stays in memory and is never written to disk.
02
Read with an AI model
Google Gemini reads the page, text and layout, and fills a fixed set of fields. Each value comes with a confidence level and a short quote from the document. Missing fields stay empty.
03
Verify with plain code
Ordinary code, no AI, recomputes the line amounts, compares the lines with the subtotal, checks subtotal plus tax against the total, and tests the dates and currency.
04
Hand over the result
You see every field with its evidence and every check with its reason, and can download or copy the result as JSON. In a real system, anything uncertain goes to a person.
Honest scope
What it is, and what it is not
It is
✓A working demo of the AI systems we build: extraction, evidence, checks and review.
✓A quick way to see how a model handles your kind of invoice layout.
✓Honest about uncertainty: low-confidence values and failed checks are shown, not hidden.
✓Private by design here: we never store your file or the result.
It is not
–A production accounting system. Nothing is posted, matched or saved.
–Perfect. Accuracy varies, and poor scans, photos and handwriting make mistakes more likely.
–A place for confidential documents. Files go to a third-party AI model to be read.
–Built for every document. The fields and checks are designed for invoices and receipts.
Need documents read on your own servers, with no third-party model? That is our private AI work.
Want this on your own documents?
We build document reading into real workflows: your fields, your checks, a review queue for the uncertain ones, and export to the systems you already use.