DoorDash Vera Data Assistant
Supports natural language queries on enterprise data using a knowledge layer and evaluation mechanisms. Task: DoorDash built Vera, a data analyst agent that answers natural language questions using a knowledge layer of table descriptions, metric definitions, and business rules. The system retrieves relevant tables, generates SQL, and validates queries against user permissions. For heavy analysis, the agent presents a plan for human approval before execution. Evaluation mechanisms gate new data sources, requiring them to pass accuracy thresholds before activation.; tools: Tool described in the source. Human role: The original summary does not fully specify human review and operational roles; actual deployment should include human verification based on task requirements.; limitations: Source is a secondary compilation of public engineering blogs; verify original DoorDash posts for exact architecture. Implementation requires a robust knowledge layer (metadata, metrics, rules) and a comprehensive evaluation suite. Human approval is mandatory for heavy analysis plans. New data sources must pass eval thresholds before being enabled. Cost and latency metrics are not provided..
Tools
Tool described in the source
Published on Agent Work Atlas:
Overview
Supports natural language queries on enterprise data using a knowledge layer and evaluation mechanisms.
Workflow
DoorDash built Vera, a data analyst agent that answers natural language questions using a knowledge layer of table descriptions, metric definitions, and business rules. The system retrieves relevant tables, generates SQL, and validates queries against user permissions. For heavy analysis, the agent presents a plan for human approval before execution. Evaluation mechanisms gate new data sources, requiring them to pass accuracy thresholds before activation.
Reported result
Vera covers over 200,000 datasets. The evaluation pass rate improved from 43% to 90% across more than 1,900 test cases. On the same underlying model, Vera scored 2.60 out of 3 compared to 2.16 for plain Claude Code, indicating the harness and knowledge layer significantly improved performance. Specific deployment scope and ongoing operational metrics beyond these evals are not detailed in the source.
Human role
The original summary does not fully specify human review and operational roles; actual deployment should include human verification based on task requirements.
Limitations
Source is a secondary compilation of public engineering blogs; verify original DoorDash posts for exact architecture. Implementation requires a robust knowledge layer (metadata, metrics, rules) and a comprehensive evaluation suite. Human approval is mandatory for heavy analysis plans. New data sources must pass eval thresholds before being enabled. Cost and latency metrics are not provided.
Try it
Build a prototype data analyst agent for a small, non-sensitive dataset. 1. Create a knowledge layer: document table schemas, key metric definitions, and 5-10 example queries with business rules. 2. Implement retrieval: ensure the agent searches this layer before generating SQL. 3. Set up evaluation: create 10 real questions with known correct answers. 4. Configure permissions: restrict queries to read-only access for the test user. 5. Add approval: require human confirmation for any query scanning more than 100MB. Run the 10 eval questions. Measure: accuracy (correct answers/total), time to answer, and whether the agent correctly refused out-of-scope queries. If accuracy <80%, refine the knowledge layer. Do not deploy to production data until evals pass 90% and human approval workflows are tested.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.