Presentation: From Models to Agents: Building Context-Aware Consumer AI at Scale at DoorDash
DoorDash ditched one-shot predictions for agentic recommendations. Here's how consumer memory and semantic IDs changed the conversion game.
Why it matters
A major consumer platform moved from static ML to stateful agents that remember user context across sessions. This is a working deployment case study showing how agents handle real production scale and measurable business impact.
The key facts
11 to knowDoorDash shifted from legacy one-shot predictions to agentic recommendation platform
Uses language-native consumer memory for context retention across sessions
RQ-VAE semantic IDs for catalog representation
Grounded search integration for relevance
Dramatic boost to relevance and conversion metrics (specific numbers not disclosed in abstract)
Speaker: Sudeep Das (DoorDash engineering)
Architecture uses language-native consumer memory (stateful context across sessions)
RQ-VAE semantic IDs for catalog representation (vector-based product encoding)
Grounded search mechanism for relevance
Platform reports dramatic improvements in relevance and conversion metrics
Published via InfoQ (technical practitioner audience)
Go to the source
InfoQ AI/MLinfoq.com
Publisher excerpt: Sudeep Das shares how DoorDash shifts from legacy one-shot predictions to an agentic recommendation platform. He discusses leveraging language-native consumer memory, RQ-VAE semantic IDs for catalog representation, and grounded search to dramatically boost relevance and conversion metrics. By…