Research Systems
A compact view of the main systems I am building. The publication list remains the complete record; this page explains the project context behind selected papers and prototypes.
Wireless Sensing
Focus: Scalable biometric and vital-sign sensing with distributed UWB MIMO radar in real indoor environments.
System contribution: RF hardware, wireless synchronization, embedded data collection, and learning-based recovery of vital-sign waveforms from mixed motion and sensing signals.
Related work: MEDUSA, MobiCom 2025; Gemini, mmWave sensing and communication.
On-Device AI
Focus: Multimodal assistants that run locally on battery-powered edge devices rather than relying on cloud inference.
System contribution: NanoMind and Virgile combine custom device prototypes, embedded software, accelerator-aware scheduling, and local vision-language inference for persistent object and face recognition.
Artifact: 3D-printed hardware prototype and demo system close to completion.
ML Systems
Focus: Efficient edge inference pipelines across heterogeneous accelerators and private device runtimes.
System contribution: EdgeFlow is the project umbrella for accelerator-aware ML systems work, including Split to Fit style model partitioning and CRANE, a direct Apple Neural Engine inference runtime that bypasses Core ML.
Related work: Split to Fit, PalmBench, CRANE, and efficient mobile LLM/LMM execution.
LLM Post-Training
Focus: Reinforcement-learning fine-tuning for LLMs that must reason, remember, and retrieve under limited context and compute budgets.
System contribution: EMBER studies budgeted evidence retention for long-horizon agents, while CPPO trains coordinated pass@K policies for diverse code reasoning attempts.
Related work: EMBER and CPPO arXiv drafts.