
DoorDash built a one-click platform that replays historical support conversations through an LLM-driven customer simulator to evaluate their support chatbot at scale, tracking KPIs like escalation rate, resolution quality, and hallucination rate; the simulated escalation rate (46%) closely tracked real production traffic (44%), validating the approach as reliable test coverage before shipping changes.
First of a blog series on the engineering behind DoorDash Assistant.

Personalization at DoorDash spans restaurants, groceries, retail, and more - but capturing true consumer intent requires moving beyond sparse engagement patterns.
