Inside Vera, DoorDash’s Data Agent

Ian Baldwin, Jacopo Himberg, Akshat Khandelwal16 Sep 2026

Problem

DoorDash staff in operations, finance, marketing and product ask data questions such as why orders dipped in California last week. The company has more than 200,000 datasets and 350 petabytes, and the upstream sources are fragmented. Earlier internal tools were either deep in one domain or broad and shallow. Coding agents with a simple SQL connector performed poorly here.

Example from the article: a request such as "Show me canceled orders by hour for the last 90 days at Arby's" has to land on the right table and the right definitions, not just run valid SQL.

Solution in brief

DoorDash built Vera, a conversational data agent whose accuracy comes mostly from the scaffolding around the model: retrieval over a vetted data model, blocked access to unverified sources, and a large eval set curated by domain owners. Pass rate went from 43% to 90% as the eval set doubled twice.

Questions this case answers
  • How to make an LLM agent answer questions over thousands of overlapping, inconsistently defined tables with answers people can trust
  • How to source and maintain a representative evaluation set for an agent when only a few experts can judge correctness
  • How to decide whether to invest in a better model or in the scaffolding around it
  • How to make a data agent return answers over fragmented company data that the business can trust

Tools and methods

  • Retrieval and ranking over a modeled data layer — one embedding model searches human-authored knowledge and modeled tables in parallel.
  • LLM judge — scores tool use, table choice and answer correctness on each eval.
  • Eval mining — a frontier model drafts evals from Slack threads and Jira tickets for owners to accept or drop.
  • /analysis planning flow — plan, then user approval, then execution for hard questions.
  • Claude Code — the baseline agent in the harness-versus-model test.
Continue reading — it's free

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