Research VLM-SSM Vision Encoders Do VLMs need Vision Transformers? Controlled comparison of state space models vs ViTs as VLM vision encoders. Toward Tokenizer-Free Chinese LMs Controlled comparison of Chinese character-embedding families under a fixed backbone; first milestone of a longer research program reframing tokenization as learned token construction. ICASSP 2024 — TransMix Cross-domain augmentation for limited-supervision foot-ulcer segmentation. ICASSP 2024 (equal first author). MLISA AIE Backend (VCK5000) Runtime-only retargeting of a PyTorch → IR → MLISA pipeline from a custom RTL NPU to AMD AI Engine cores on a VCK5000 FPGA. Adaptive Chess Tutor Locally-deployable dual-LoRA TinyLlama chess tutor; original CSE 537 course project (Spring 2025) plus a 47-evaluation reproduction (May 2026) characterizing what holds up and what doesn't. Industry AI-Powered Beamforming Antenna 1st Place + Best Presentation at Inventec Hackathon 2023 — repurposing existing laptop WiFi antennas as sensors via AI, with zero hardware modifications. Stylus Trajectory CNN Lightweight CNN (10–50K params) for stylus-trajectory regression from capacitive sensors, deployed in the panel IC of a top IC design house. Sub-2-pixel accuracy, ~50% error reduction over the prior rule-based algorithm.
