Satish Chandra

My work on angelic nondeterminism, Angelic Debugging, and SemFix helped establish logical foundations for fault localization and automated program repair. SemFix received the ICSE Most Influential Paper Award in 2023. Earlier work on scalable interprocedural analysis was transferred into IBM’s commercial software-analysis technology.

I led work on applying machine learning to software development at Facebook, including Getafix and SapFix for automated program repair, Aroma and neural code search for code recommendation and search, and predictive test selection. Several of these systems were deployed in production and influenced subsequent work in industry and research.

At Google, I led work on adapting frontier language models to enterprise software engineering, including G4G (Gemini for Google), trained on internal code, diffs, documentation, and review history, and Passerine, an agentic system for issue resolution. I now lead research on coding agents at Meta, focusing on their quality, correctness, and efficiency.

I study how AI systems for software engineering should be evaluated on realistic tasks. My recent work includes REAP, RubberDuckBench, and benchmarks for code reasoning and agentic repair derived from production software activity, with a focus on whether benchmark performance reflects real developer work.