Ecommerce ops · Production

DoorDash reimagines hyper-personalization by layering LLMs over classic ML systems

The problem

Classic personalization systems built on collaborative filtering, embeddings, and deep learning could not meet customers at their moment of need or incorporate world knowledge about new inventory, while human-centric product-attribute extraction and editorial merchandising could not scale to millions of individual consumers.

First attempt

Static personalization approaches — including matrix factorization, LDA, wide-plus-deep models, and two-tower embeddings — learned only from long-term in-app engagement and could not react fast enough to short-lived, high-context moments like Black Friday or new merchant inventory with no engagement history.

Workflow diagram · grounded in source
1
LLM product attribute extraction
ai_action
“we started with vanilla LLMs. They get you somewhere. It's the 80 of the 80/20. Because we cannot actually sell something for which the allergen was wrongly extracted because someone might actually die, so, basically, we need to be very …”
2
Agentic small-merchant enrichment
ai_action
“with that text, if you go and search on Google, normally, it finds a lot of information that's very relevant, even if it's highly abbreviated. Then from all the information we gather from Google, and also, we can go to the merchant's web…”
3
Human-in-the-loop validation
human_review
“We have been able to solve it from a purely human-driven process because those items are so hard, to something that's very much AI-driven with human in the loop. The story is human in the loop.”
4
Consumer narrative profile generation
ai_action
“The latest trend has been like, let's just represent consumer profiles in just plain old English… These consumer profiles that we talk about, we group them into different memory blocks about dietary habits, household information, categor…”
5
Offline carousel blueprint generation
ai_action
“for every single consumer, every week, for different use cases, we essentially ask the LLMs again saying that like, what should we show to this consumer… Most of the content generation about what to showcase to every single consumer is g…”
6
Real-time intent blending via DoorDash Brain
integration
“you have this service of DoorDash Brain, which understands everything about the consumer, which hosts also your profiles and serves them, in addition to that, it also does the blending of your real-time intent”
7
Embedding retrieval and ranking
ai_action
“your two-tower embedding models or whether it is your MTML ranker models, don't throw them away. Still, basically, you need to use them to make sure that they encode a lot more information about specific item preferences that you essenti…”
8
GEPA prompt optimization feedback loop
feedback_loop
“We use GEPA right now within DSPy. GEPA is Genetic-Pareto, which is one of the ways in which you can optimize these compound AI systems… we use LLM-as-a-judge as well… if you're able to collect human annotated feedback, especially just t…”
Reported outcome

DoorDash reduced a product-attribute extraction task from 28 days to 2 days using LLMs with fine-tuning, RAG, and agentic processes, and built a hyper-personalization system that generates per-consumer carousel blueprints offline and blends real-time intent signals online to serve moment-aware, individualized experiences.

Reported metrics
Product attribute extraction cycle time28 days to 2 days
Carousel ideas generated per consumer40 to 50
Embedding retrievals per homepage request20 to 30
Reported stack
LLMsRAGDSPyGEPAMTMLGoogle
Source
https://www.infoq.com/presentations/llm-personalization/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash reduced a product-attribute extraction task from 28 days to 2 days using LLMs with fine-tuning, RAG, and agentic processes, and built a hyper-personalization system that generates per-consumer carousel bluepr…

What tools did this team use?

LLMs, RAG, DSPy, GEPA, MTML, Google.

What results were reported?

Product attribute extraction cycle time: 28 days to 2 days; Carousel ideas generated per consumer: 40 to 50; Embedding retrievals per homepage request: 20 to 30 (source-reported, not independently verified).

What failed first in this deployment?

Static personalization approaches — including matrix factorization, LDA, wide-plus-deep models, and two-tower embeddings — learned only from long-term in-app engagement and could not react fast enough to short-lived,…

How is this ecommerce ops AI workflow structured?

LLM product attribute extraction → Agentic small-merchant enrichment → Human-in-the-loop validation → Consumer narrative profile generation → Offline carousel blueprint generation → Real-time intent blending via DoorDash Brain → Embedding retrieval and ranking → GEPA prompt optimization feedback loop.

Related ecommerce ops cases