I build machine learning systems and tools for exploring the physical world—from generative motion and sensing to electromagnetic and neural simulation.
2026 Interactive simulation
A real-time 3D driving experiment coupling a fruit-fly connectome subgraph to simplified spiking neurons, interactive traffic, and live neural activity.
Engineered sensing and control; the model has not been trained to drive.
2026 Connectome exploration
An interactive MaleCNS connectome playground with 166,606 classified neurons. Move light, food, and sound sources to explore sensory responses and a virtual fly's motion.
Simplified LIF neurons and engineered flight control; not a validated model of fly behavior.
2026 Scientific computing
A CUDA FDTD solver that runs Tidy3D Simulation objects on a local NVIDIA GPU, built for differential validation against the reference solver.
Single-GPU, fp32 time stepping; grid and material generation come from the Tidy3D client.
Generating expressive avatars requires coordinated body, hand, and facial motion, while available datasets often annotate only some of these components. T2M-X learns from partially annotated data using separate VQ-VAE experts for each component and a multi-indexing GPT with a motion consistency loss. The model combines these representations to generate coordinated whole-body motion from text.
This project generates dance motion conditioned on music while allowing text-based edits such as “move your right arm up.” Motion and text are mapped into a shared latent space, with a motion codebook conditioned on music features—including genre, spectrograms, and tempo—to produce and edit avatar movements.
A wrist-worn gesture recognition system using inertial measurement unit (IMU) and photoplethysmography (PPG) signals, enabling interaction beyond a camera’s field of view. A modified MobileNet trained on 0.4 million samples achieved average recall of 96% and precision of 94% across gesture types. Incorporating handcrafted signal features reduced power consumption by 35% relative to the compared methods.
Demo · Patent · Yahoo coverage · Nomtek coverage · TNW coverage
A human activity recognition system for aging in place using mmWave radar instead of RGB cameras. Four deep learning models were compared, with the best model reaching 91.87% test accuracy. The prototype connects Raspberry Pi devices to AWS IoT for data collection and processing, with an AWS Lambda threshold for fall alerts.
Project website · Demo · Presentation · Code
A GAN-based approach to designing dielectric resonator antennas for the 60 GHz band. A simulator learns the relationship between antenna geometry and the S11 response, while a generator proposes geometric patterns conditioned on a target S11 spectrum. The system reduces reliance on manual trial-and-error design.
