jeremyefarr/po-pilot

PO Pilot is a system that aims to automate corporate purchasing from start to finish. Instead of employees filing purchase requests through separate emails, chat messages, or forms, PO Pilot lets them ask for what they need in plain language through Slack, Microsoft Teams, email, or a dashboard, and one automated agent handles all of it. It is built on top of Hermes Agent, a project from Nous Research, and was made for a contest built around that framework. The system checks every request against a company's purchasing rules before any money moves, so approvals and rejections happen automatically based on policy rather than manual review. When a purchase is approved, it issues a Stripe virtual card that is locked to a specific vendor, capped at a set spending limit, and usable only once. After the purchase, PO Pilot connects to a company's bank data through Plaid to match the original request, the receipt, and the actual bank charge down to the penny, automating the reconciliation work that would otherwise be done by hand. Behind the scenes it combines a language model, reasoning through Hermes Agent, with a sandboxed component called NemoClaw that enforces the purchasing policy at the moment a request is made. It is designed to run on NVIDIA DGX Spark hardware on premises, so company data does not leave the building. The backend is built with Django and Celery, the dashboard with Next.js, and Playwright automates parts of the checkout process in the browser. The README includes a demo file that plays through three example requests at once: one small purchase gets automatically approved and a card is issued, one falls into a blocked spending category and is rejected outright, and one exceeds a spending threshold and gets escalated to a manager based on the company's reporting structure. Together these show how the system is meant to route different kinds of purchase requests without a human needing to process each one manually.