Invoices into validated data.
Upload an invoice and get structured fields. Every number is re-computed in code and reviewed next to the line it was read from.
In plain words
You upload an invoice and get its fields as data: vendor, dates, amounts, line items. Before you see them, code re-does the arithmetic. Anything that does not add up is marked with the reason, next to the line of the document it was read from.
How it works
Upload a PDF or paste text
A sample invoice opens by default, one where the lines do not add up.
See what was read and what failed
Every field has a status: ok, missing or inconsistent, with the reason in plain words.
Fix it and export
Editing a value re-runs the checks at once. The checked result is copied as JSON.
The problem
Whatever reads a document can be wrong in a way that looks fine: a dropped line, a swapped digit, a total taken from the wrong row.
What we built
Extraction and checking are separate. After the fields are read, code verifies the arithmetic: quantity times price, line items against the subtotal, subtotal plus tax against the total. Anything that fails is marked with a plain reason, and a person fixes it on a review screen that re-checks as they type.
How we know it works
| Sample invoices, three layouts, two currencies | 8 |
|---|---|
| Fields read correctly on the samples | 64 of 64 |
| Deliberately broken invoices caught | 2 of 2 |
| Automated tests | 34 |
The samples and the rules were written together, so this is a regression check, not a benchmark. Known limits are listed in the repository.
- Built with
- TypeScript, Express, PDF text extraction
- Model
- Optional; a rule-based reader runs by default
- Output
- JSON, after a human review step
- Good fit for
- Accounting intake, logistics paperwork, back office
Under the hood
- PDFText layer, lines rebuilt from positions
- ReadRules by default; a model is optional
- CheckArithmetic, dates and sources, in code
- ReviewA person fixes what is flagged
- JSONExported after the review
Scanned pages without a text layer are rejected; there is no OCR. The model path is tested against a stubbed endpoint and has not been run on a live provider yet.
Decisions and why
Reading and checking are separate
Arithmetic does not depend on the reader behaving, so the same checks run on rules and on a model.
Rules first, model optional
The default reader is deterministic and needs no key, so the demo is free and repeatable.
A person stays in the loop
A wrong vendor name can still add up. The review screen exists for what arithmetic cannot catch.
Limits are written down
Whole-number prices, wrapped descriptions and scanned pages are listed in the repository as known gaps.