dbt Summit
All on demand sessions
On demand session

From prompt to PR: sandbox-validated dbt changes before human review

Careem's AI agent writes and validates dbt models end to end before a human ever reviews the PR.

Recorded at dbt Summit · 2026Breakout session archive
01

Overview

Breakout session · 2026

A five-person data warehouse team managing 100+ concurrent requests and 5,000+ active ETLs moved beyond "let Claude generate SQL, have engineers review it." Implements a four-layer validation architecture (SQL compilation, dbt tests, Airflow orchestration, OpenMetadata lineage) that must pass before human review; 45% of new models now carry an "ai-generated" tag.

What you'll take away

  1. 01

    Confidence in AI-generated code comes from end-to-end validation, not the code generator — Careem's agent compiles, tests, and materializes every change in a sandbox before a human reviews it.

  2. 02

    A spec-locking step makes the agent ask clarifying questions about grain, metric definitions, and filters upfront, closing off a common failure mode where "revenue" could silently mean three different things.

  3. 03

    The results: metric delivery time dropped from two-to-three weeks to hours, about 90% of new models now come from this framework, and severity incidents have fallen close to zero.

Return to all sessions
02

Session speakers

02 speakers

  • Niyazulla Khan Pathan

    Careem

  • Akash Srivastava

    Careem