Inspiration
AI shopping agents can recommend and submit purchases, but they should not have unrestricted payment authority. Till adds a policy checkpoint between an agent’s order proposal and PayPal checkout.
What it does
Shoppers describe what they need and set spending caps, exclusions, a review threshold, and a delivery address. Till uses Gemini to match the request to an item in its demo catalog and assess whether it fits. Deterministic checks validate shopper intent, catalog identity, quantity, address, total, spending limits, and duplicate orders. The model can flag an item or request human review, but cannot override a hard block or authorize payment.
How we built it
Till is built with Python, FastAPI, SQLite, Gemini, and the PayPal Orders API in Sandbox. After the policy decision, Till creates and reads the PayPal order; the buyer authorizes it before capture. Till verifies and deduplicates PayPal webhooks, then checks that capture amount and currency match before marking an order complete. Two MCP tools let external agents submit proposals and retrieve decisions without accessing PayPal credentials or payment controls.
What we learned
The main design challenge was separating AI judgment, deterministic policy, shopper approval, and PayPal payment state. The model assesses product fit, while code enforces spending limits, item integrity, idempotency, and payment confirmation.
Current scope
Shopping uses a demo catalog, and payments run through PayPal Sandbox.
