Research
Software is too important to be unreliable or insecure. It underpins our digital world and increasingly controls physical systems. Yet software is increasingly being created and changed faster than people can review it, accelerated by AI-powered development. My research pursues two complementary goals: building intelligent tools that can automatically audit software, find and repair bugs and vulnerabilities, and establishing the methods needed to assure that these tools and the broader AI-enabled systems are reliable and secure.
AI for Software Engineering
Software development is moving beyond what human review alone can keep pace with. We develop AI-enabled methods that audit code and software artifacts, identify defects and vulnerabilities, trace them to their causes, generate repairs, and validate that proposed fixes are correct. Our goal is to turn AI into a dependable engineering partner that helps developers build and maintain reliable, secure software at scale.
Representative papers
PatchGuru: Patch Oracle Inference from Natural Language Artifacts with Large Language ModelsarXiv · 2026Memory-Efficient Large Language Models for Program Repair with Semantic-Guided Patch GenerationICSE · 2026Athena: Vulnerability-Affected Library Identification via Knowledge Graph CompletionEMNLP · 2026Assurance for AI-Enabled Software
AI is increasingly used both to create software and as a core component of software systems, from coding assistants to autonomous agents. Failures and vulnerabilities in these systems can propagate into generated code, automated decisions, and real-world actions. We develop rigorous evaluations, realistic benchmarks, and practical defences to understand these risks and improve the reliability and security of AI-assisted development and AI-powered software.
Representative papers
Optimizing Spectral Signature in Code Backdoor DetectionICSME · 2026How Do Semantically Equivalent Code Transformations Impact Membership Inference on LLMs for Code?ICSE · 2026Can LLMs Reason About Program Semantics? A Comprehensive Evaluation of LLMs on Formal Specification InferenceACL · 2025