About me
I am a PhD candidate in Computer Science and Engineering at the University of Michigan, advised by Professor Kang G. Shin and Professor Ke Sun. I expect to graduate in May 2028.
My research builds resource-efficient multimodal AI systems through task-aware inference and serving. I work on vision-language model (VLM) inference, edge-cloud orchestration, and serving, focusing computation on the evidence a task actually needs to reduce latency and cost.
Current research
- Task-Aware VLM Serving for Procedural Assistance moves inference preparation ahead of anticipated task-check requests with speculative prefill around a vLLM server. In single-stream replay, it reduces median time-to-first-token from 2.34 s to 0.37 s (p95 from 4.1 s to 1.5 s), and task-aware evidence selection improves F1 over uniform frame sampling.
- Thermal-Kitchens is an RGB-thermal benchmark evaluating VLM cross-modal reasoning about hidden physical states, built from 29 hours of synchronized video and temperature measurements across 49 cooking episodes.
Recent work
- InvAgent is an agentic edge-cloud system for food-inventory monitoring that treats the task as sparse-evidence inference. An edge context builder and a bounded agentic workflow focus VLM analysis on amount-evidence windows, reducing VLM input tokens by 9.9x versus 1-FPS whole-video inference while increasing amount-aware inventory F1.
My earlier work includes UWB-PTrac, a single-anchor in-vehicle phone-localization system published at ACM/IEEE SenSys 2026, and Light Auditor, a power side-channel approach to detecting private-data leakage through IoT covert channels, published at ACM SenSys 2022.
I received B.S. degrees in Computer Science and Mathematics from the College of William & Mary, where I worked with Professor Gang Zhou.
I am available for a full-time position from May 1 to August 31, 2027, and will return to my PhD afterward.