Robinhood data-platform engineer Chetas Joshi describes an AI-assisted reviewer for data-pipeline pull requests. His reported 60% automated merge share is for this bounded class of PRs, not the company’s entire codebase.
Give facts and judgments different owners
- Scripts check schema contracts, table owners and sensitive-data labels, recorded downstream consumers, and performance against staged workload tests.
- A repository-reading agent looks for risks those checks may miss. It can inspect code and access grants but cannot query live catalogs or claim that an absent repository match proves no consumer exists.
- Merge needs passing checks and an approving review. Sensitive-data grants, meaning-changing migrations and uncertain verdicts go to a person even when the agent sounds confident.
Follow a schema change across the boundary
- An example adds a column and widens another from integer to big integer. The registered consumers pass deterministic checks.
- The agent finds an unregistered job in code that writes the widened value into an integer column elsewhere. A human confirms the job is active and updates its destination table before merging.
- This catches a gap in recorded lineage; dashboards and ad-hoc queries outside the repository remain unseen.
Measure both misses and review burden
- Human replies inform a weekly manual review-rule update; the system does not automatically retrain or rewrite its own rules.
- Joshi reports about 60% of eligible PRs merging without human review, 75% less manual-review effort, about 132 engineering hours saved monthly and roughly 30% fewer related incidents. These are his internal observations, not controlled causal estimates.
- He reports an approximately 2% miss rate as of August 31, 2026, but no clean single false-escalation rate. A system that escalates everything saves little reviewer time.
- His proposed unified reviewer with live-check tools is not the deployed architecture. Today the scripts and agent run separately and cannot see each other’s runtime state.
“None found is not the same as none exist.” — Chetas Joshi