Ecommerce ops · Production

DoorDash builds LLM-powered consumer, merchant, and item profiles to enable explainable personalization at scale

The problem

DoorDash's entity representations relied on opaque embedding vectors that were impossible to explain to consumers or debug, limiting the platform's ability to build interpretable personalization and recommendation features.

Workflow diagram · grounded in source
1
Input data gathering
integration
“This foundational stage involves gathering, cleaning, and structuring the vast array of data points necessary to generate comprehensive profiles of consumers, merchants, and items”
2
LLM profile generation
ai_action
“These engineered features are then fed into an LLM, which is prompted to generate a detailed, human-readable profile. This is where the "magic" happens; the LLM synthesizes the various data points into a coherent and insightful descripti…”
3
Offline model comparison
validation
“we then compare profile outputs across different model providers and between reasoning and non-reasoning models using sampled but representative cases. We have discovered that employing a reasoning model with a medium reasoning effort du…”
4
Batch production pipeline
integration
“Both consumer and merchant profile pipelines now run in Fabricator as batch jobs”
5
Post-processing and storage
output
“This step focuses on appending extra fields and attributes to the LLM-generated profiles for various use cases and purposes, allowing us to accommodate data that doesn't require LLM inference”
6
LLM-as-judge quality evaluation
feedback_loop
“use LLM-as-a-judge to automatically score profile quality using source data, generated profile, and a rubric, enabling rapid offline iterations”
Reported outcome

LLM-generated narrative profiles enable explainable recommendations, editable consumer preferences, and batch processing at scale, cutting per-profile spend by roughly 50% and processing millions of records within an hour.

Reported metrics
Per-profile batch spendroughly 50%
Batch throughputmillions of records within an hour on a modest cluster size
Profile generation time vs manualfraction of the time it would take to do so manually
Reported stack
LLMsPySparkFabricatorAmazon S3PortkeySQLSpark
Source
https://careersatdoordash.com/blog/doordash-profile-generation-llms-understanding-consumers-merchants-and-items/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LLM-generated narrative profiles enable explainable recommendations, editable consumer preferences, and batch processing at scale, cutting per-profile spend by roughly 50% and processing millions of records within an…

What tools did this team use?

LLMs, PySpark, Fabricator, Amazon S3, Portkey, SQL, Spark.

What results were reported?

Per-profile batch spend: roughly 50%; Batch throughput: millions of records within an hour on a modest cluster size; Profile generation time vs manual: fraction of the time it would take to do so manually (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Input data gathering → LLM profile generation → Offline model comparison → Batch production pipeline → Post-processing and storage → LLM-as-judge quality evaluation.

Related ecommerce ops cases