Chetas Joshi

Chetas Joshi helped build the automated reviewer for Robinhood’s data pipelines. As a senior software engineer on the data platform team, he designed and built distributed systems that bring data into Robinhood’s data lake and process it for analysis. When coding assistants began opening pull requests faster than the team could review them, he and his colleagues automated much of that review while reserving decisions about access to sensitive data for engineers. Chetas reported that, as of August 31, 2026, about 60% of the team’s data pipeline pull requests merged without human review, saving roughly 132 engineering hours a month.

Biomedical imaging and data platform engineering

For his 2015 master’s thesis in electrical and computer engineering at Purdue, Chetas used functional MRI during a breath-holding task to measure how readily blood vessels in the brain responded. He compared scans taken before, during and after the football season in high school players who showed no symptoms of injury. Across three seasons, he found a significant first-half decline in that response only in the season when players experienced the most head impacts.

Before Robinhood, Chetas built data platforms at Oracle Data Cloud and Airtable and was a founding data platform engineer at Rubrik. At Rubrik, one fault could trigger a cascade of alerts, leaving the engineer on call to work out which one pointed to the cause. Chetas and five colleagues patented a method for alert dependency discovery that compared incoming alerts with past patterns to identify which warnings tended to follow others. Those relationships could help an engineer decide where to investigate first.

Finding dependencies that pipeline checks miss

At Robinhood, the team kept scripts responsible for questions with an authoritative answer, such as whether a new column existed in the upstream schema. It gave an AI agent access to repository files so it could trace how other jobs used a changed column. In one case, every dependency in the platform’s records could handle a wider integer. The agent found an unregistered job that copied the value into a column with a narrower range. An engineer confirmed the job was running in production, and the team widened the destination column before merging the change.

That case also explains a design choice Chetas came to question. The scripts and the agent ran independently. The agent knew which checks the scripts were supposed to perform, but it could neither see their results nor run a live check itself. Chetas described a proposed revision that would let the agent request a specific check when its review depended on a fact about the running system.

Chetas judges the reviewer by two kinds of mistakes. An unsafe approval can break a pipeline. An unnecessary escalation makes an engineer spend time on a change the system could have cleared. His team asked engineers to explain why they agreed or disagreed with a review, and the engineer on call used those explanations to revise the rules each week. One missed dependency became a specific rule: when a privacy classification moves a column into a restricted copy of a table, check for jobs that still read it from the open copy.