I am a Senior Principal Researcher at Microsoft M365 Research, where I am the tech lead and architect for end-to-end inference efficiency of agentic workflows: workflow characterization, orchestrator–inference engine co-optimization, KV-cache management and novel methods for efficient inference.
My recent AI work includes task-specific LLM compression (LLM-Sieve: 20–75% parameter reduction at ≤1–5% accuracy loss) and prospection-guided agentic memory (EMNLP 2026), which introduced the public MemoryQuest benchmark.
Machine learning has run through my whole career. I started with neural-network pruning and feature selection (1997–99), and my best-known systems work on indoor localization (EZ, Zee, Centaur) is built on Bayesian inference, particle filters and custom optimization. Over 20 years at Microsoft Research and Bosch I have turned research into shipped systems, including Azure ML labeling (~10× cost reduction) and the next-generation Xbox controller protocol (5× lower latency).
I have published 50+ peer-reviewed papers with 9,500+ citations, hold 30+ US patents, and received the ACM Test of Time Award in 2022. I have mentored researchers who are now faculty at UIUC, Cornell, NYU, Georgia Tech and UC Irvine, and served on the FCC Technological Advisory Council AI Working Group (2022–23).
