Checks a supplier invoice against what was ordered and what actually arrived. AI reads the document and pairs differently worded lines; rules in code decide whether it can be auto-approved. The demo records decisions, not payments.
An upload becomes a saved decision
FastAPI stores the upload and queues its ID. A LangGraph worker reads the document, looks up the supplier, purchase order, and goods receipts, matches the lines, and checks for duplicates. Each step and the final decision are saved so a reviewer can inspect what happened.
The hosted system uses AWS Lambda, SQS, and S3, with PostgreSQL on Neon and the interface on Vercel. Local development swaps in an in-process queue and PostgreSQL file storage behind the same interfaces.
Code owns the approval
Exact item codes and descriptions are matched first. For unresolved lines, the model can choose only from supplied purchase-order candidates, or choose none. Invalid choices, ambiguity, and missing records require review.
The rules check a 2% upper price tolerance, delivered quantities, duplicates, currency, and a known supplier. Auto-approval also requires a total strictly below the configured limit. Failures are shown as plain review reasons; the model never computes the final approval status.
Permissions and bounded usage
Each visitor receives an isolated demo company for 24 hours. A reviewer cannot decide an invoice they uploaded, rejection needs a reason, and final decisions cannot be changed. The API enforces these rules, not just the visible controls.
Live AI requests are capped at five per visitor and 100 across visitors each day. Usage is reserved atomically in PostgreSQL before each call, and provider retries are disabled. Reaching a limit routes the invoice to review; saved examples still work without model calls.
Checks and known limits
Normal CI exercises 14 labelled synthetic decision cases with scripted model responses. All 14 keep their expected decisions and review reasons, with zero wrong auto-approvals. Any wrong auto-approval fails the check; these are software tests, not AI benchmarks.
Separate real-model runs on 11 November 2026 scored 40 of 45 exact fields across three invoice images, and 84 of 84 across four generated invoices. These tiny samples are repeatable evidence, not a representative accuracy estimate. Decision evaluations start from known fields and do not measure document reading.
AI can still misread a document or pair a line incorrectly, and rules can only check the values they receive. This is not a payment system or production approval authority. Demo roles are not authenticated identities; uploads should be fictional and non-sensitive.