I work on machine learning and the systems that run it: protein structure prediction, fast image generation, and language models that reason well or run on small devices. These days I mostly teach computer science, one-on-one.
I studied Electronic Science and Technology at Shanghai Jiao Tong University, then received my M.S. in ECE from Purdue University, where I worked with Prof. Felix Xiaozhu Lin on video analytics for IoT cameras. I then joined the CS Ph.D. program at Georgia Tech, advised by Prof. Alexey Tumanov, and am currently on leave.
Since then I have worked in industry research, first as a research scientist at an AI drug discovery startup, and later at a GPU company on research collaborations with research institutes.
Predicting a protein's 3D shape from its amino-acid sequence, and what mutations do to it. OmegaFold makes high-resolution predictions from a single sequence, without the multiple sequence alignments that AlphaFold2 depends on. FastFold makes AlphaFold training and inference faster on GPU clusters. In a study of influenza neuraminidase, lab measurements of more than 6,000 mutations agreed well with fitness inferred by a protein language model.
Diffusion models turn noise into an image over many small steps. InstaFlow uses rectified flow to straighten that path, which makes it possible to distill Stable Diffusion into a generator that needs only one step.
Training models to reason better, in math and across text and images, and measuring how small language models actually run on phones and other edge devices.
Also: PhoneLM: An Efficient and Capable Small Language Model Family through Principled Pre-training (arXiv 2024) and On the Perception Bottleneck of VLMs for Chart Understanding (arXiv 2025).
Video analytics on energy-constrained IoT cameras, deep learning for identifying wireless interference, and visual analytics for hyperspectral images.
Also: Deep Learning for Frame Error Prediction using a DARPA Spectrum Collaboration Challenge (SC2) Dataset (IEEE Networking Letters 2021) and Efficient Training of Deep Classifiers for Wireless Source Identification using Test SNR Estimates (IEEE Wireless Communications Letters 2020).
Alongside research, I have taught more than 3,000 hours of one-on-one lessons in university computer science, math and statistics. That covers 150+ courses at 40+ universities, for students in Australia, the UK, Hong Kong, Singapore and North America.
I start from where a student is stuck, explain down to the mechanism, and move on only once they can explain it back.