DoorDash builds a multi-agent internal AI platform for unified data and operations intelligence
The problem
DoorDash's knowledge was distributed across experimentation platforms, metrics hubs, dashboards, wikis, and Slack, requiring significant context-switching to answer complex business questions.
First attempt
Existing self-service tools were sub-optimal because they assumed users already knew which data sources to query and how to interpret them correctly; skillset gaps and the risk of misinterpretation limited their usefulness for critical analyses.
Workflow diagram · grounded in source
1
User submits business query
Trigger
A user submits a complex business question, such as investigating a drop in conversions.
▾ source quote
“Investigate the drop in conversions in the Midwest last week”
2
Multistage contextual retrieval
Ai action
A multistage search engine combines best-match-25 keyword search and dense semantic search with reciprocal rank fusion re-ranking to retrieve relevant context for the agent.
▾ source quote
“our engine addresses this by using an algorithm that combines traditional best-match-25 keyword search with dense semantic search, followed by a sophisticated re-ranker using reciprocal rank fusion, or RRF”
3
Schema-aware SQL generation
Ai action
The DescribeTable AI tool provides the agent with compact, engine-agnostic column definitions enriched with pre-cached example values to improve filtering accuracy.
▾ source quote
“we use our DescribeTable AI tool with pre-cached examples. This tool provides the agent with compact, engine-agnostic column definitions. Crucially, it enriches this schema information with example values for each column that are pre-cached in an in-memory store. This significantly …”
4
Statistical query validation
Validation
A multi-stage Zero-Data Statistical Query Validation and Autocorrection process checks query correctness and performance without exposing sensitive data to the AI model.
▾ source quote
“rigorous, multi-stage validation process we call Zero-Data Statistical Query Validation and Autocorrection. This includes automated linting for code style and markdown enforcement, but its core is an EXPLAIN-based check for query correctness and performance against engines like Snowflake and Trino. …”
5
LLM-as-judge quality evaluation
Feedback loop
An automated LLM-as-judge framework grades each agent response for accuracy and uses DeepEval to measure faithfulness and contextual relevance.
▾ source quote
“automated LLM-as-judge evaluation framework. For a platform intended to guide high-stakes business decisions, "good enough" isn't an option. This framework systematically runs predefined question-and-answer scenarios against our agents. An LLM judge grades each response for accuracy and provides a detailed …”
6
Answer delivered in Slack or IDE
Output
Agents are invoked directly within Slack or Cursor, delivering answers without requiring context-switching.
▾ source quote
“An analyst investigating a trend can pull data directly into a Slack conversation, or an engineer can generate boilerplate code without leaving their editor”
Reported outcome
DoorDash's agentic platform now enables operators to get trustworthy, evidence-backed answers in seconds, democratizing data access for business leaders and operations managers without requiring them to write SQL.
Reported metrics
Time to answeranswer in seconds, not hours
Decision-making and execution speeddramatically accelerating decision-making and execution across the company
Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.
DoorDash's agentic platform now enables operators to get trustworthy, evidence-backed answers in seconds, democratizing data access for business leaders and operations managers without requiring them to write SQL.
What tools did this team use?
LangGraph, Google Docs, Google Sheets, Slack, Cursor, Jira, DeepEval, MCP, A2A, vector database.
What results were reported?
Time to answer: answer in seconds, not hours; Decision-making and execution speed: dramatically accelerating decision-making and execution across the company (source-reported, not independently verified).
What failed first in this deployment?
Existing self-service tools were sub-optimal because they assumed users already knew which data sources to query and how to interpret them correctly; skillset gaps and the risk of misinterpretation limited their usefu…
How is this back office ops AI workflow structured?
User submits business query → Multistage contextual retrieval → Schema-aware SQL generation → Statistical query validation → LLM-as-judge quality evaluation → Answer delivered in Slack or IDE.
This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.