Open sourceReleased

invoice-extractor

Turns photos and PDFs of invoices into rows ready for the books: Claude reads them, accounting rules check every number, and what does not add up goes to a person.

Invoice and receipt reading in .NET 10 with Claude Haiku 5.5: deterministic validation after the model, a queue in PostgreSQL, human review beside the original, and export to Excel and CSV.

Role:: Design and development
Year:: 2026

View Code

false auto-approvals in the evals: 0/42
planted errors sent to review, none approved: 7/7
accuracy on total, date, tax id and number, with the retry: 100%
model cost per document: $0.00065

75 seconds with Claude Haiku 5.5: three of the built-in examples, two approved by the rules, one corrected by a person. On-screen text in Spanish.

The review screen: a photographed receipt on the left and, on the right, the fields read, with the total flagged in red because subtotal plus tax does not give it — What does not add up goes to review: the original alongside, the failing field flagged.

A batch of three documents: a PDF and a scan approved automatically, and a photo that needs review — A batch of three: two pass every rule and are approved on their own.

Dashboard with 3 documents, 67% approved automatically, the total model cost and the average cost per document — The dashboard: how many are approved on their own, and what reading each one costs.

The problem

Many small businesses still type their invoices into the books by hand. A language model can read them, but it is sometimes confidently wrong, and a misread invoice in the books is worse than one left out. The goal was a reader you can trust: one that approves only what it can check, and sends the rest to a person with what they need to decide quickly.

Constraints

Nothing is approved on the model's word alone: every reading is checked by rules
Tilted photos, poor scans, thermal receipts and PDFs, in Spanish and English
Tax rules per country: Nicaragua's RUC and VAT, Mexico's RFC and VAT, and a generic profile
Exact amounts end to end: decimals, never floating point
A visible cost per document and a daily spending cap

How it works

Each file is checked on its own (real format, size, pages) and recognised by its hash: uploading the same one twice does not cost twice.
A PostgreSQL queue with FOR UPDATE SKIP LOCKED and a lease hands documents to the workers: each one is read exactly once, even if a worker crashes.
Claude Haiku 5.5 reads a digital PDF's text or the image of a scan or photo, with structured output against a JSON Schema. Amounts travel as text and become decimals without ever being a double.
Deterministic rules decide: the lines add up to the subtotal, subtotal plus tax gives the total, the rate exists in that country and the tax id has its format. A failing reading is read once more with the errors as a hint; if it still fails, it goes to human review.
A Next.js web app, in Spanish and English, shows each document's status live, the review beside the original and a cost dashboard, and exports to Excel and CSV.

Key decisions

The model proposes, the rules decide

Arithmetic and formats can be checked without AI. A reading that does not add up is never approved on its own, however confident it looks, and the model can also say it is unsure of a field.

Trade-off: Rules catch inconsistencies, not every error: a misread value that still adds up would pass. The evals count those cases, and so far there are none.

One retry, with the errors as a hint

Most misreadings are a digit or a letter. Told which rule failed, the model fixes it: on the hard set, tax-id accuracy goes from 91.7% to 100%.

Trade-off: About 50% more cost on hard documents, and the ones that are really wrong are paid for twice before they reach a person.

A queue in PostgreSQL, not a broker

SKIP LOCKED and a lease give exactly-once processing with the database the app already has. A test with four workers and a hundred documents proves it.

Trade-off: It polls the database every few seconds and has a throughput ceiling, far above what an accounting office uploads.

Results

On 42 synthetic eval documents, with Claude Haiku 5.5 and the retry: 100% accuracy on total, date, tax id and number, and no false auto-approval.
The 7 documents with deliberately planted errors end in human review in every run.
Without the retry, the hard set misses the tax-id pass mark (91.7% against 95%). That failure is published alongside the successes.
USD 0.00065 per document on the standard set and USD 0.00097 on the hard one, retries included.
A Playwright end-to-end test drives the real app (upload, review, approval and Excel) on every push, with recorded model answers: no API key and no cost.

.NET 10
C#
PostgreSQL
Aspire
Claude Haiku 5.5
Next.js
TypeScript
Playwright

All projects

Demo

Screenshots