Systems Research / On-Device & Edge AI
Building intelligence that runs locally.
Built from the ground upI am a systems researcher building hardware, software, and models for on-device and edge AI. I focus on running LLMs and VLMs locally on mobile and wearable devices. My work spans efficient inference, RL post-training, and wireless sensing.
I obtained my Ph.D. from the University of Wisconsin-Madison, advised by Prof. Suman Banerjee.
Over the past several years, I have built and deployed MIMO radar systems in indoor and hospital environments, worked with Dr. Eugene Chai and Dr. Karthikeyan Sundaresan at NEC Laboratories America, and explored silicon and ML acceleration for wearables at Meta Reality Labs. These experiences taught me that practical AI cannot be designed one layer at a time: sensing hardware, accelerators, runtimes, models, and deployment constraints must be considered together.
This perspective now guides Re-Mind, my work on privacy-preserving, on-device memory for personal assistants. Re-Mind slides
AI must not be locked inside corporate black boxes. Free intelligence that people themselves can run, study, modify, reproduce, share, and trust.
02 /Research in practice
From local multimodal inference
to agents and sensing systems.
Showing all 6 selected works.
Tiny but Mighty: A Software-Hardware Co-Design Approach for Efficient Multimodal Inference on Battery-Powered Small Devices
A battery-powered multimodal inference system that coordinates model modules, heterogeneous accelerators, and shared memory to run a self-contained assistant locally.
Read the research
PalmBench: A Comprehensive Benchmark of Compressed Large Language Models on Mobile Platforms
An automated benchmark of compressed mobile LLMs that brings resource use, latency, throughput, answer quality, and harmful outputs into one evaluation.
Read the research
Virgile: A Multimodal Visual Memory Assistant with Persistent Object and Face Recognition on Edge Devices
A wearable visual-memory assistant built on NanoMind hardware, combining local vision-language inference with persistent object and face recognition.
Explore the system
Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning
A code-reasoning post-training method that coordinates alternative plans and a shared solver, rewarding valid strategy sets only when execution confirms success.
Read the research
Scalable Biometric Sensing in the Wild through Distributed MIMO Radars
A distributed UWB MIMO radar system that combines wireless synchronization, multiple spatial views, and learned signal separation for contactless vital-sign sensing.
Read the research
EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents
A learned memory policy that preserves source-backed evidence before future questions are known, helping long-horizon agents work within a fixed evidence budget.
Read the research