Chinmay Gandhi

Chinmay Gandhi helped build tools at Meta that could start investigating a production failure as soon as an alert fired. By November 2026, he was a senior engineering manager at Meta. His work ranged from turning engineers’ diagnostic playbooks into code to testing whether an AI investigator had found the cause of an incident and proposed a fix someone could use.

Earlier, while studying at Syracuse University, Chinmay adapted ResourceSpace, an open-source system for managing digital collections, for the university’s Collage photo-sharing project. He later served as director of software engineering at Adara, where he worked on Magellan, a self-service platform for digital advertisers. He also co-invented a method for controlling access to pooled data that let each contributor specify which operations others could perform on its data.

Automating an incident investigation

Chinmay co-authored the research on DrP, Meta’s platform for running diagnostic playbooks. Engineers wrote the steps for investigating a failure as reusable programs called analyzers. An alert could trigger an analyzer automatically. If its checks pointed to a problem in a service owned by another team, it could pass its findings to that team’s analyzer. In December 2025, Chinmay and his co-authors reported that DrP ran 50,000 analyses a day across more than 300 Meta teams and had reduced average incident resolution time by 20 percent.

Weighing the evidence for a diagnosis

With Vlad Tsvang and Patrick Walsh, Chinmay developed the Agent Feed, where specialist agents investigating the same incident could read and build on one another’s findings. A coordinating agent ranked competing explanations by checking them against independent sources. Chinmay also co-authored Opsmate Evals, which assessed an investigation once engineers knew how the incident had ended. The team used incident conversations, code fixes and retrospective reports to reconstruct what happened. Model-based graders could then assess three questions separately. Did the agent identify the cause? Did the evidence it gathered support that conclusion? Could an engineer act on its proposed fix? Agent Feed and Opsmate Evals

Chinmay worked with Swaapnika Motapothula and Vlad Tsvang on an AI teammate shared by an on-call rotation. He illustrated the design with a slow checkout. A latency alert, customer tickets and a recent change to a payment timeout could all point to the same failure, yet an engineer still had to piece those clues together. The rotation’s agent could connect them, check the suspected cause against earlier incidents and draft a rollback for review. Its memory carried the team’s corrections into the next shift. The team allowed the agent to close routine alerts on its own once it had a track record on that action and mistakes had limited consequences and were easy to detect. An engineer still decided whether to roll back a production change.