Probie

Probie puts AI clerks on probation in accounts payable. AP teams lose hours to invoice exceptions: amounts that don’t match the purchase order, short shipments, duplicate bills, and payment requests from fake vendors. AI could clear many of them, but one wrong payment is real money lost. So finance teams either avoid AI or trust it everywhere, and neither works. Probie treats each Meta model like a new hire. Every model starts with no authority and reviews invoice exceptions against a written finance policy. Its decisions are scored against what a finance team would have done. When a model proves itself on one kind of exception, it’s cleared to handle that kind. The moment it pays a scammer, it loses that clearance. Companies can then automate only the decisions a model has actually earned, and keep people on the rest. The test is built to be hard. Invoices arrive as clean PDFs, scans, and phone photos from different vendors. Vendor messages come as real-looking emails, including spoofed lookalike domains and prompt-injection attempts hidden in the text. Models use tools to check purchase orders and vendor records, and every promotion or demotion is backed by statistics, not luck. A hard spend cap keeps costs predictable. Built solo in Python on Meta models, deployed with Docker on Render, open source under MIT. Live demo: https://probie.onrender.com