Ecommerce ops · Production

DoorDash builds a GenAI-powered personalized homepage carousel system

The problem

DoorDash's original heuristic-based content system with around 300 curated carousels lacked sufficient concept diversity, produced overly broad and impersonal recommendations, and suffered from suboptimal knowledge graph tagging that matched stores to irrelevant carousels or omitted relevant ones.

Workflow diagram · grounded in source
1
Consumer profile and day-part input
trigger
“Takes as input consumer profile and part of day — for example, breakfast or lunch — then uses the LLM to generate carousel titles and metadata”
2
LLM carousel title and metadata generation
ai_action
“We use a sophisticated LLM-powered system to generate personalized carousel titles for the homepage, driven by comprehensive consumer profiles”
3
Carousel embedding generation
ai_action
“Converts the generated carousel titles and metadata into text embeddings”
4
LLM-as-jury content moderation
validation
“This begins by prompting three different LLMs with our review criteria and then subjecting their independent decisions to a veto process. If any juror LLMs find a title to be in violation, it is automatically blocked. This moderation pro…”
5
GPU-based exact KNN store retrieval
ai_action
“Instead of a typical approximate nearest neighbor approach, we perform an exact KNN search on GPU”
6
Store ranking and carousel serving
output
“we leverage our existing store ranker to determine the order in which the stores will be displayed within each carousel. This model is optimized around engagement signals such as click-through rate or conversion rate”
Reported outcome

The GenAI carousel system improved store retrieval precision@10 from 68% to 85%, delivered double-digit click rate improvement, and drove improving conversion rates and homepage relevance metrics.

Reported metrics
Store retrieval precision@10 (baseline)68%
Store retrieval precision@10 (improved)85%
Content moderation recall on bad titles95%
Click rate improvementdouble-digit click rate improvement
Show all 6 reported metrics
store retrieval precision@10 (baseline)68%
store retrieval precision@10 (improved)85%
content moderation recall on bad titles95%
click rate improvementdouble-digit click rate improvement
conversion ratesimproving
homepage relevance metricsimproving
Reported stack
LLMsSpark
Source
https://careersatdoordash.com/blog/doordashs-next-generation-homepage-genai/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The GenAI carousel system improved store retrieval precision@10 from 68% to 85%, delivered double-digit click rate improvement, and drove improving conversion rates and homepage relevance metrics.

What tools did this team use?

LLMs, Spark.

What results were reported?

Store retrieval precision@10 (baseline): 68%; Store retrieval precision@10 (improved): 85%; Content moderation recall on bad titles: 95%; Click rate improvement: double-digit click rate improvement (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Consumer profile and day-part input → LLM carousel title and metadata generation → Carousel embedding generation → LLM-as-jury content moderation → GPU-based exact KNN store retrieval → Store ranking and carousel serving.

Related ecommerce ops cases