Chetas Joshi

About Chetas Joshi

Chetas Joshi is an engineer on Robinhood's data platform team, where he designs and builds the distributed data ingestion and processing frameworks behind the company's data lake. His team owns review of changes to production data pipelines: schema changes, pipeline code, configuration and access grants. In an AI Engineer talk, he explains how Robinhood reviews those pull requests now that AI assistants open them faster than people can read them.

Notable Positions

On Scripts for Facts, Agents for Judgment

Anything a source of truth can answer, such as the schema registry, catalog tags, registered lineage or a staging performance test, is checked by a deterministic script that blocks the PR on failure. The agent's rules tell it what those checks already cover, so it spends its judgment on what they can't see, such as downstream jobs missing from lineage.

On Flag, Never Clear

Grants touching sensitive data, migrations that change what data means, and verdicts the agent isn't confident in always reach a person. The agent may only surface the risk and its reasoning. See Human-in-the-Loop.

On Measuring the Reviewer

Robinhood tracks misses (bad PRs approved, about 2% as of August 31, 2026) and false escalations (fine PRs sent to a person). Misses become new rules, which the on-call engineer updates by hand each week from reviewer feedback. See Agent Evaluation.

Key Quotes

• "Pull requests to our production data pipelines now arrive at machine speed, but reviewing them safely still runs at human speed." (on the problem)
• "None found is not the same as none exist." (on the agent's limits)
• "A gate that escalates everything is not actually saving anyone any time." (on false escalations)

Related Reading

• Human-in-the-Loop - Where people stay in charge of agent decisions
• Agent Evaluation - Measuring whether agents do the job