I'm Michael — a computer-science & math student at Harvey Mudd, working in the Zhuang group on the statistical physics of deep neural networks and the structural heterogeneity of supercooled water: mean-field theory, large MD simulations, order-parameter embeddings.
Outside the lab I build tools at the seam of atoms and algorithms — agentic spectroscopy reasoning, sparse autoencoders on RL-trained models, GNN pipelines for MS/MS fragmentation. I care about systems you can hold in your head, and results that survive being looked at twice.
Read more about this work here
Relevant coursework: Statistical Mechanics, Probability, Statistics, Differential Equations, Graph Theory
Awards:
Click on any project to learn more
An agent that reads NMR, IR, and MS spectra, runs six specialist ML models across four evidence streams, and returns ranked structures — showing its reasoning at every step. Runs as a CLI, a web GUI, or an MCP server inside Claude or Cursor.
Is liquid water one continuum or a mixture of two local structures? I let unsupervised clustering decide on MD trajectories, then validated it against an observable the clustering never saw. Two structurally distinct populations emerge on their own.
Can sparse autoencoders trained on a chain of PPO checkpoints tell you what RL fine-tuning changes inside a model? A three-axis scorecard separates reward-invariant reorganization from reward-graded causal load.

atoms, agents, photos, and what's in between.