AI Invoice Processing System
Extraction, confidence thresholds and a review queue
Python · Claude API · PostgreSQL · FastAPI · Docker · n8n
What it does
Ingests invoices from a mailbox or folder, extracts structured data with per-field confidence, validates the arithmetic, and routes anything below your thresholds to a review queue with the source region highlighted. Corrections feed back as labelled examples.
Who it is for
Finance teams processing hundreds to low thousands of invoices a month, and consultants implementing this for clients.
Why the review queue is the product#
Extraction accuracy is the part everyone benchmarks. Review time per flagged document is the part that determines whether the system saves anyone anything. This package optimises both, and the second one harder.
Requirements
Docker, PostgreSQL 15+, a vision-capable language model API key.
Questions
What accuracy should I expect?
On clean digital PDFs from recurring suppliers, high. On handwritten or poorly scanned documents, low — those go to the review queue, which is the design working as intended rather than failing.
Is it tied to one automation tool?
No. Orchestration templates are provided for more than one platform, and the extraction service is a plain HTTP API you can call from anything.
Can I use this for client work?
Yes, under the single-team commercial licence. Reselling it as your own product is not permitted.
Changelog
- v1.4.01 Aug 2026
Source region highlighting in the review queue.
- v1.3.011 May 2026
IBAN checksum validation; multi-currency parsing.
- v1.2.028 Feb 2026
Correction feedback loop.
How this was built
AI Invoice Processing System
An invoice pipeline that extracts structured data, scores its own confidence per field, and routes anything uncertain to a person — because 95% accuracy on invoices is a finance problem.