Namaste 🙏 · hello 👋

I'm Abhijith Sharma. I audit tool-using LLM agents.

I'm a PhD student in Machine Learning at MBZUAI in Abu Dhabi, advised by Salem Lahlou and Nils Lukas. I work on auditing, testing and verification of AI agents: training auditors that interrogate tool-using LLM agents the way real users would, and fuzzing multi-agent systems for the failures that only appear when agents, tools and memories interact.

Previously, I was a research associate at the University of Waterloo with AVL, where I built a real-time adversarial testing framework for autonomous vehicles under Sebastian Fischmeister and Nasser Azad. Before that I was a MITACS Accelerate intern at TrojAI, working on physical adversarial attacks and defenses for vision models under Apurva Narayan, which became my master's thesis at UBC.

I did my bachelor's at the College of Engineering, Pune, graduating with the gold medal, and spent three of those years on the control subsystem of COEP-SAT-2, a student satellite. There was also a short stint in decision analytics at ZS Associates.

Outside of work I play squash, swim, and read history.