01 / AI SECURITY
Making AI systems easier to trust.
LLM applications can fail without a useful trace. I work on methods that make those failures easier to isolate, and combine language models with symbolic engines to check answers.
As tech lead and science lead at AWS AI Security, I set technical direction across research and engineering. My current work spans testing LLM-integrated applications, agent security, and privacy.
See how it works & related research
- The question is a region Under-specified questions to rule engines define regions, not points. An LLM maps the stated facts to a symbolic region; concolic execution covers every reachable outcome, and Z3 verifies the resulting cited partition. When it is too large, the system asks one targeted follow-up. NeSy 2026.
- Delta debugging for LLM-integrated systems LLM applications fail opaquely: no execution paths, no stack traces, non-deterministic outputs. We wrap each input segment in a verifiable semantic marker, have the model cite markers in its reasoning, and use delta debugging to shrink hundreds of segments to the few that cause the failure. ICSE-SEIP 2026.
- Advances and open problems in federated learning The foundational survey of federated learning, covering privacy, robustness, and systems challenges. Foundations and Trends in Machine Learning, 2021.