When GenAI Meets Personalization: Powering DoorDash's next-generation homepage experience

Yuxiang Wang, Yefei Wang, Sicong Fang, Siyao Xiao, Rui Hu, Anish Manne, Aruj Padbidri, Yang Yu, Di Li2 Dec 2025

Problem

DoorDash's homepage showed about 300 hand-curated carousels, such as "breakfast burritos" or "salads". Each user was matched to carousels by comparing tags from past orders with carousel tags in a food knowledge graph (FKG). Customers said 300 concepts were too few, concepts like "Salads" were too broad and impersonal, and poor tagging put irrelevant stores into carousels or left relevant ones out. Every new carousel also needed manual creation and tagging.

Example from the article: the same title "Hearty wraps" is turned into paneer and chicken tikka wraps for a user with Northern Indian preferences, and into buffalo chicken and Caesar wraps for a user with American preferences.

Solution in brief

An LLM reads a consumer profile and the part of the day, and writes carousel titles plus structured metadata that mirrors merchant profile fields. Title and metadata are embedded and matched to merchant and dish embeddings with exact KNN on GPU. Three LLMs vote on moderation, and a block re-ranking step blends the engagement ranker with embedding similarity.

Questions this case answers
  • How to generate personalized content themes for each user with an LLM and fill them with matching items from a catalog at scale
  • How to moderate millions of generated texts without human review while keeping recall of violations high
  • How to combine an engagement-trained ranker with a new relevance score that the ranker has no training data for
  • How to replace a fixed set of hand-tagged homepage carousels with carousels generated for each user?

Tools and methods

  • LLM carousel generation — titles and metadata from the consumer profile
  • Text embeddings and exact KNN on GPU — retrieve stores and dishes
  • LLM-as-jury moderation — three LLMs with a veto rule filter bad titles
  • Block re-ranking — blends the store ranker with similarity
  • Apache Spark — builds prompts and runs batch LLM requests
Continue reading — it's free

Google or a one-time email code. No password.