Download the latest PDF resume.
InvAgent: Agentic Edge-Cloud Food Inventory System (July 2025-June 2026)
First author; manuscript under review. Formulated food-inventory monitoring as sparse-evidence inference: brief observations reveal item use and remaining amount, while whole-video VLM analysis spends computation on irrelevant frames. Designed an edge context builder that summarizes item interactions and a bounded agentic workflow that selects amount-evidence windows and revisits unresolved items. Across approximately 39 hours from seven households, reduced VLM input tokens 9.9x versus 1-FPS whole-video inference while increasing amount-aware inventory F1 from 66.6% to 68.9%.
EventHOI: Event-Driven Hand-Object Interaction Logging (August 2025-March 2026)
Coauthor; manuscript under review. Developed the temporal-localization evaluation and characterized frame-rate, resolution, and compression tradeoffs. Across 50 hours and more than 19,000 verified interactions, EventHOI reduced RGB frames by 50-65% and bandwidth by 2.7x with a 1.0-2.0 percentage-point accuracy loss.
Task-Aware VLM Serving for Procedural Assistance (June 2026-present)
Ongoing research; preliminary results. Implemented a speculative-prefill prototype around a vLLM server, using experimental prefix caching to move inference preparation ahead of anticipated task-check requests. In single-stream recorded-video replay, reduced median time-to-first-token from 2.34 s to 0.37 s and p95 from 4.1 s to 1.5 s, trading additional compute for lower request-time latency. Built task-aware evidence selection; preliminary evaluation on 147 labeled examples of correct and mistaken execution showed a 3-percentage-point F1 improvement over uniform frame sampling.
Thermal-Kitchens: RGB-Thermal Perception and Reasoning (March 2026-present)
Co-first author; ongoing research. Contributing to an RGB-thermal benchmark evaluating VLM cross-modal reasoning about hidden physical states, using 29 hours of synchronized video and temperature measurements across 49 cooking episodes. Benchmark results show that paired RGB-thermal input does not consistently improve VLM performance; answer accuracy was 81.8% when thermal evidence was correctly grounded versus 12.9% otherwise.
UWB-PTrac: UWB-Based In-Vehicle Phone Localization (August 2023-July 2025)
First author; ACM/IEEE SenSys 2026. Formulated in-cabin phone localization under the constraint of existing UWB keyless-entry hardware: a single, single-antenna anchor insufficient for conventional triangulation. Developed a single-anchor phone-localization approach that resolves ambiguous channel impulse response signatures through optimized anchor placement or a lightweight RF shield. City-driving tests achieved 96% driver-vs-other-seat classification accuracy with optimized anchor placement and 91% with the RF shield.
Kailai Cui, Kaylee Yaxuan Li, Chaoyu Zhang, Peizhou Huang, Hao Chen, Jianzhong Zhang, Ke Sun, Kang G. Shin, "Agentic Edge-Cloud Orchestration for Resource-Efficient Daily Food Inventory Monitoring with Egocentric Video." Manuscript under review.
Le Zhang, Kailai Cui, Hao Chen, Vlad Roznyatovskiy, Jianzhong Zhang, Anhong Guo, Kang G. Shin, Ke Sun, "EventHOI: Event-Driven Energy-Efficient Hand-Object Interaction Logging on Smartglasses." Manuscript under review.
Kailai Cui, Ke Sun, Kang G. Shin, "UWB-Based Localization of Smartphones inside a Vehicle to Prevent Distracted Driving," ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys), 2026.
Woosub Jung, Kailai Cui, Kenneth Koltermann, Junjie Wang, Chunsheng Xin, Gang Zhou, "Light Auditor: Power Measurement can tell Private Data Leakage through IoT Covert Channels," ACM SenSys, 2022.