System Design
How DoorDash Handles Real-Time Order State at Scale
DoorDash reduced subtle support-chatbot hallucinations by replacing prompt-only fixes with simulation, binary evals, and a cleaner case-state layer.
1. DoorDash's LLM support chatbot was hallucinating not dramatically, but subtly. It would misread a delivery status field and confidently suggest a refund policy that didn't exist. The fix wasn't a better prompt. It was a better system.
Expand thread
2. They built a simulation + eval flywheel: → An LLM plays the customer using real historical transcripts as behavioral seeds, frustrated, patient, escalating realistically → Another LLM judges each conversation against a specific binary policy check.
3. → 200+ conversations, results in under 5 minutes. 11 iterations. 90% hallucination reduction.
4. The counterintuitive fix: give the chatbot less information, not more. Raw order history, tool calls, delivery logs all dumped into context was creating noise. They added a "case state" layer, a clean intermediate summary of only what matters. More context ≠ better reasoning.
5. What DoorDash built is what real LLM engineering looks like simulation, calibrated eval, architectural layers that control what the model actually sees. That's the kind of system I want to build. Not exploiting LLMs. Actually engineering with them.