Presentation: From Models to Agents: Building Context-Aware Consumer AI at Scale at DoorDash
| Source: InfoQ AI/ML
Tags: DoorDash, RQ-VAE, recommendation systems, agentic AI, consumer AI, semantic search
DoorDash's ML head Sudeep Das describes how the company rebuilt its recommendation engine around LLM-driven consumer memory and RQ-VAE semantic IDs, replacing one-shot predictions with an agentic search pipeline that improved relevance and conversion across grocery, convenience, and retail verticals.
Details
Recorded at QCon AI Boston 2026, this presentation details DoorDash's shift from legacy prediction models to an agentic recommendation architecture. The core components: language-native consumer memory (storing user context as LLM-readable representations), RQ-VAE semantic IDs for catalog items (hierarchical embeddings enabling semantic search across millions of SKUs), and grounded agentic search that retrieves results with explicit factual backing. Sudeep Das leads ML and AI for DoorDash's New Business Verticals — covering grocery, convenience, alcohol, and retail. The talk covers how DoorDash moved from one-shot relevance predictions to multi-step agent workflows that can iterate and refine results before returning them to users. The InfoQ summary is sparse on specifics — no latency numbers, no A/B test deltas, no architecture diagrams. The presentation recording is available, and the real technical depth lives there rather than in this article. For teams building recommendation systems at e-commerce scale, the semantic ID approach and consumer memory design are worth exploring.