LangGraph · structured output · vector match · human in the loop
Invoices in, validated data out. The model reads; tested code decides.
Pick a synthetic invoice or upload your own. The agent extracts the fields into a strict schema, matches the vendor against a vendor master, runs 13 validation rules and either approves the invoice or stops for a human.
Loading…
How it works
- The model reads. An LLM (or the free demo extractor) fills a strict Pydantic
Invoice; output that breaks the schema is retried once with the error fed back. - Tested code decides. 13 plain-Python rules: arithmetic, tax rate, dates, currency, vendor match (a small vector store), tax id, duplicates, PO balance, goods receipt (three-way match), lookalike characters and injected instructions.
- A person handles exceptions. Any failed rule pauses the graph with LangGraph's
interrupt(); the reviewer approves, rejects, corrects the fields, or adds the vendor and re-checks. Corrections go back through every rule. - Everything is recorded. Each node's timing and outcome, every decision in an audit log, and the same agent is exposed as MCP tools for Claude Desktop. No approve tool: the model never approves.
1 · Invoices
Try an attack
PDFs with a text layer (exported from accounting software, a browser or Word), or a photo: photos are read by OCR in your browser (Tesseract.js) and only the text is sent; OCR results always go to a person. Do not upload invoices with personal data to the public demo.
Queue
Vendors you added
2 · Graph path
Pick a sample to run the graph.