Building the AI Layer for a Food Delivery Support Bot🚀

A regional food delivery platform needed a customer-facing AI agent that could handle live order queries in Arabic and English — without hallucinating status, misrouting frustrated customers, or breaking under injection attempts.

My role was the AI behavior layer: system prompt architecture, decision logic, and the test suite that kept it honest. I designed an 11-gate decision tree where routing is determined by order status × elapsed time — not customer tone. A frustrated message doesn't trigger escalation. A VENDOR_ACCEPTED order delayed 20 minutes does.

The test suite ran 41 versions across 10 categories: routing thresholds, language switching, ETA edge cases, multi-turn flows, and injection resistance.

Result: live in production for over a month, 98% customer satisfaction rating from the support team.