Research Systems
A compact view of three current research lines. The publication list remains the complete record; this page explains the project context behind selected papers, prototypes, and ongoing work.
RL Finetuning for Small Models
Focus: Improving small language models under limited context, memory, and compute budgets.
Ongoing project: Reinforcement-learning fine-tuning, supervised fine-tuning, and agentic-memory policies that decide what to write, retain, and recall for downstream solvers.
Representative work: EMBER studies budgeted evidence retention for long-horizon agents; CPPO trains coordinated pass@K policies for diverse code reasoning attempts.
Wireless Human Sensing
Focus: In-the-wild pupillometry and cognitive sensing with a glasses-mounted 60 GHz mmWave radar and front-facing illumination context.
Ongoing project: mmPupil estimates pupil dynamics with radar, uses the front-facing camera to model light-driven pupil changes, and subtracts that component before workload inference.
Status: Under review. The draft is not publicly linked.
Representative work: MEDUSA, LiveTag, and Gemini frame the broader wireless human sensing direction.
On-Device AI
Focus: Privacy-preserving cognitive assistants and multimodal AI devices that run locally rather than relying on cloud inference.
System contribution: Re-Mind, NanoMind, and Virgile combine custom device prototypes, embedded software, accelerator-aware scheduling, local vision-language inference, and real-world episodic memory.
Artifact: 3D-printed hardware prototype and demo system close to completion, backed by runtime and benchmarking work such as PalmBench, Tiny but Mighty, and CRANE.