DoorDash Cuts Grocery Reorder Latency from 16 to 2 Seconds with AI Agents

| Source: InfoQ AI/ML

Tags: DoorDash, Ask DoorDash, batch inference, LLM latency, InfoQ, agent platform

DoorDash cut its AI grocery reorder latency from about 16 seconds to 2 by precomputing recurring purchases with batch LLM inference, which doubled shopping list views and lifted order rate about 34%.

Details

InfoQ reports that DoorDash stopped rebuilding grocery reorder lists on each request and now precomputes them. Batch LLM inference over 90 days of purchase history and profile signals predicts recurring purchases, and deterministic validation checks results before they are stored. At serving time, each item is resolved against the live catalog: first the original store item, then a catalog product ID, then a search term. The change took latency from about 16 seconds to 2, doubled shopping list views and raised the flow's order rate by roughly 34%. It runs inside Ask DoorDash on a shared agent platform that handles orchestration, memory, model access, tracing, evaluation and rollout.