Robinhood data-platform engineer Chetas Joshi said a hybrid reviewer now merges 60% of data pipeline PRs with no human review, per AI Engineer. Before that, CI plus human review gated every schema change, query and grant. Joshi said the demo covered lake ingestion tables, the schema registry, catalog and lineage tool, with no remote attacker in scope.
Scripts prove facts. The agent judges code.
Layer 1 scripts check the registry for contracts, the catalog for owners and PII tags, registered lineage for blast radius, and staging at production volume for freshness SLAs, Joshi said. On the example PR, a new column plus an integer to bigint widen passed all four checks.
The agent then walked repo code and found an unregistered reader that wrote the widened column into a table still typed as integer, Joshi said. It scored the risk high and escalated with the full chain for a person to confirm. Joshi cautioned that "None found is not the same as none exist" for repo search.
Feedback is weekly and manual. An on-call engineer reads human replies and updates rules by hand, Joshi said. As of , the miss rate was about 2%, manual review effort was down about 75%, time saved was about 132 engineering hours a month, and incidents tied to these PRs were down about 30%. Grants touching sensitive data, intent-changing migrations and low-confidence verdicts still escalate by design.
Layers 1 and 2 now run separate with no shared runtime state, and the agent cannot call live checks. The platform team proposal for one classifier entry point with checks as tools is not built, and the weekly hand update continues.
