Lead Scientist for Amazon Q Developer (f.k.a CodeWhisperer) -- data-filtering, pre-training, fine-tuning of code generation model; post processing with re-inforcement learning based on execution and static analysis feedback, incorporating project context using lexical/semantic/structural retrievals to enhance response accuracy/relevance, enterprise customization, reducing hallucinations using program analysis, unit test generation, automated bug fixes, evaluations based on real benchmarks and A/B testing. Results include general availability of Q Developer, enterprise customization, doubling acceptance rate of code suggestions, increase in number of accepted chars per user resulting in enhanced retention of users.

1. Architect of Piranha, a tool for automatically deleting code related to stale feature flags and designed an automated workflow for Piranha that generates diffs periodically. Lead cross org efforts advocating the deployment of Piranha for stale feature flag cleanup. The system is now used across the entire mobile organization at Uber as part of Engineering Fundamentals Code Quality initiative and has deleted approximately 5000 flags (∼300KLoC) in the mobile apps implemented in Java, Objective-C and Swift for Android and iOS platforms. 2. Led the design, implementation and deployment of tracking unit test coverage, static analysis issues and custom business-specific metrics across all Uber codebases (xx MLoC of Java, Go, Swift, JS, Kotlin, Scala) on SonarQube. (Uber Infra Business Impact Award Q3 2021, Uber Developer Platform Velocity Award Q4 2021)3. Rolled out workflows to enable detection of data races in Java and Go backend services code, resulting in 1000s of fixed data races (Uber Infra Tech Award Q2 2021, PLDI 2022 paper)4. Led the design and implementation of systems to manage flaky tests and built tooling to help reproduce flaky tests locally. (Uber DevPlatform Velocity Award Q3 2020)

The primary focus of my research was on automatic test generation for detecting concurrency bugs. We built test generators that analyze sequential execution traces and construct bug-revealing multithreaded tests to expose deadlocks, races, atomicity violations and assertion violations in Java libraries. Our (artifact evaluated) research prototypes demonstrated the success of our techniques in revealing critical defects in well tested libraries including the latest versions of JDK. This research was generously supported by Google. The secondary focus of my research was on designing efficient profiling tools for large Java applications. A key result included the design of a profiler that can efficiently identify the memory usage across various control flow paths in the program. This data can be leveraged by analysis tools to detect unnecessary memory allocations, thereby reducing the pressure on the garbage collector. The publication detailing our technique and the experimental results received the ACM SIGSOFT Distinguished paper award. Apart from this, I worked on designing other miscellaneous program analysis tools. The relevant details can be found in the reports listed in my google scholar page.

Member of the core analysis team at Coverity. The static analysis engine is used to analyze Java, C#, C/C++ programs to detect different bugs. My key responsibility included the design and implementation of checkers for reporting concurrency bugs. Other responsibilities/successes included scaling the static analysis engine using parallelism, improving the runtime and memory overhead of the dynamic analysis engine and unification of the analysis framework so as to use one analysis framework to report defects for different programming languages.

Designed and implemented a smart editor that enabled physicians to file medical reports efficiently.