Principled LLMs for reliable systems software.

I develop program analysis techniques for finding bugs and vulnerabilities in complex systems software. A central line of my Ph.D. research is The Hitchhiker's Guide to Program Analysis, a three-part research series on practical bug detection for the Linux kernel and related systems, with the first two parts published at OOPSLA and ASE. This line of work motivates my broader interest in how LLMs can help developers detect, localize, and reason about failures in large software systems.

As AI agents become part of software workflows, they introduce new failure modes, security risks, and engineering challenges. My new research series, The Gap Between User Intent and Code Implementation, studies the intent that code leaves implicit, the assumptions LLMs add when reasoning about it, and how program analysis and human input can help check and clarify those assumptions.

We discuss a more principled approach to integrating LLMs (i.e., mostly harmless LLMs) with static analysis for practical bug analysis in large codebases.

BugLens guides LLMs through structured reasoning steps to post-refine taint-style static analysis reports in the Linux kernel.

LLift integrates LLMs with static analysis to improve practical bug detection in large codebases such as the Linux kernel.