From Prompt to PR: Shipping AI-Written Data Models Without Losing Trust

Data modeling is the bottleneck in most analytics orgs: a handful of engineers are the only path to production, and analysts wait weeks for a metric. The obvious fix — let an LLM write the SQL and have an engineer review the PR — fails, because reviewers pattern-match on code; they don’t execute it, and one bad join cascades across thousands of pipelines.

This talk shows how a five-person team flipped the model: trust comes from execution, not review. An AI agent, grounded in a metadata knowledge graph, turns an analyst’s request into a data-model PR that must clear a four-layer sandbox — compile, tests, real materialization in an isolated environment, and lineage checks — before a human makes one approve/reject call. Isolation is enforced structurally, not by policy.

The result: time-to-metric fell from weeks to under 24 hours, 90% of new tables are agent-authored, zero incidents across 5,000+ pipelines. You’ll leave with the architecture, the guardrail blueprint, and an honest map of where AI still needs a human.