DoorDash built a GenAI-powered personalized carousel system for their homepage that generates unique store recommendations for each user. The system uses LLMs to analyze customer profiles and create customized carousel titles with metadata, then retrieves relevant stores using embedding-based search. This replaced their previous 300-carousel heuristic system, addressing issues with concept diversity and relevance. The pipeline includes carousel generation, embedding creation, content moderation using LLM-as-jury, store retrieval via GPU-accelerated KNN search, and ranking that balances engagement with relevance. A/B tests in San Francisco and Manhattan showed double-digit click rate improvements and better conversion rates.
Table of contents
Overcoming existing system limitationsLooking at the big pictureGenerating carousels and titlesExpanding queries with metadataModerating contentRetrieving stores and dishesDetermining ranking and presentationEvaluation and experiment resultsFuture workAcknowledgementsSort: