Agent reasoning and evaluation · Post-training · Synthetic data
I’m a senior applied scientist at Microsoft working on AI agents for Copilot. I use synthetic data and human feedback to build and validate evaluations, diagnose reasoning and tool-use failures, and guide post-training.
I completed a Ph.D. in Human-Computer Interaction at Carnegie Mellon University. I’ve also spent time at MIT, the Allen Institute for Artificial Intelligence (AI2), and Conservation X Labs, where I helped develop team-formation tools for global environmental innovation contests.
I build benchmarks and synthetic tasks to evaluate agent reasoning and tool use. I develop agentic quality graders and validate evaluation signals through human judgments and controlled variations in output quality. I also study how tool design shapes planning, task success, and computational cost.
I build pipelines for synthetic training data and quality grading. I contribute data and reward signals to supervised fine-tuning and reinforcement learning to improve agent reasoning, tool use, and grounded responses.
I study how agents can resist manipulation and protect sensitive information. My work includes pipelines that generate diverse adversarial tests from everyday user requests to measure attack resistance alongside successful completion of legitimate tasks, and HypoVeil for privacy-aware collaboration.
I build systems that help people decompose problems, discover ideas across domains, and synthesize scientific literature. Projects such as BioSpark, Synergi, and Threddy combine human framing and interpretation with AI retrieval, abstraction, and synthesis.
* Equal contribution
Earlier version: CMU Agent Workshop · 2025.
Best Paper Honorable Mention
Best Paper Award