Revenue Leakage Defender· Open-source prototype
Catch overbilling before you pay it.
Upload a supplier invoice and its framework contract. You get every discrepancy and a draft dispute email.
Invoice checks are slow, so they get sampled.
Checking every supplier invoice against its framework contract by hand takes too long, so teams check a sample. Rate errors, wrong line totals and non-compliant payment terms slip through as overbilling.
Customer: me. I wanted to see how far GenAI can go in a sales and finance workflow before a human has to step in.
How it works
Read the documents
PDFs are parsed in memory. An unreadable file goes straight to the report instead of being guessed at.
Invoice Framework contractText in memoryBoth files are read in memory. Nothing is written to disk. Extract invoice and contract
Two LLM calls run in parallel and return structured data, validated against Pydantic schemas.
Invoice LLM call Invoice dataFramework contract LLM call Contract dataTwo LLM calls run in parallel. Each result is validated against a Pydantic schema. Check the maths without AI
Line totals, VAT, rate caps, payment terms and validity are checked in plain Python with exact decimals. Every amount must appear in the PDF text.
Line totals VAT Rate caps Payment terms ValidityPlain Python, exact decimals. No LLM involved. Ask the LLM only when needed
When the invoice breaks a contract term, the LLM weighs whether a clause justifies it. Arithmetic errors never go to the LLM.
Arithmetic error Flagged, no LLMContract term broken LLM weighs the clauseArithmetic errors are flagged directly. The LLM only judges contract terms. Report and draft
Each invoice ends in one of three outcomes: approved, discrepancies found or review required. A dispute email is drafted for a person to review and send.
Approved Discrepancies found Review required Dispute draftThree possible outcomes, plus a draft email that a person reviews before sending.
Built to be trusted
Nothing saved to disk
Uploaded PDFs are read in memory only.
Fails closed
Incomplete or unverifiable data means review required, never approved.
A human decides
The dispute email is a read-only draft. A person reviews it before anything is sent.
Swappable LLM provider
One provider interface: Gemini, a mock for tests, or another model through environment variables.
Observable
Live progress for each pipeline step in the UI, plus LangSmith tracing.
Tested offline
51 tests that need no real LLM. Fake models cover the worst cases: one justifies every charge, one times out, one invents amounts.
Inside the app
Upload the invoice and the contract. 
Every discrepancy, with severity and reasoning. 
A dispute draft, ready for human review.
Built with
- Python
- FastAPI
- LangGraph
- Pydantic
- pdfplumber
- Gemini
- LangSmith
Runs locally today. The design maps onto AWS: API Gateway and Lambda, Bedrock, S3, DynamoDB and SES.