Co-founder and CTO, Erised AIPhD Biophysics, UC Berkeley
I build AI that doesn’t just predict, but explains how complex physical and biological systems actually work.
NeurIPS 2026 oral, sole author NeurIPS 2025 Nature Biomedical Engineering 2026 Thesis
As the brain works, the blood flowing through it changes its magnetism by a hair. An MRI scanner can follow that change across the whole cortex, and from the pattern it records we can now tell what a person is listening to, and something of what they are imagining. A thought, in other words, can be read from the magnetism of blood.
Underneath that fact lies the question that first drew me to science. Atoms are cold, interchangeable, and indifferent; yet from assemblies of them arise cells, sentences, and the minds that read them. How that happens, how meaning arises from matter, has been my question ever since. It took me, after two years in the business school at Peking University, to philosophy and physics at Vanderbilt; then to the physics of living matter in Vijay Balasubramanian’s laboratory at the University of Pennsylvania; and finally to the biophysics program at UC Berkeley, where I joined Jack Gallant’s lab and found a second intellectual home at the Redwood Center for Theoretical Neuroscience with Bruno Olshausen, who later served on my thesis committee.
At Berkeley the question became an empirical one, and that scanner was the instrument. Its recordings are noisy, high-dimensional, and only indirectly related to the activity of neurons; the method for reading them came from the Redwood Center: sparse coding, which assigns each learned axis a single meaning. With it, I used machine learning to decipher how abstract concepts such as number, space, and time are represented across the human cortex. A second project reversed the direction: drawing on the understanding such models provide, we built closed-loop deep brain stimulation that decodes a patient’s movement from the brain itself and adjusts therapy for Parkinson’s disease in real time.
Both projects point to the same conclusion. The principal constraint on scientific progress is seldom the quality of our models. We consistently overestimate our own ingenuity: the hypotheses we are able to write down form a measure-zero subset of what is possible, and genuine progress has come, far more often than from new ideas, from new and better measurements made in the environments where the systems of interest actually operate. This is a hopeful conclusion, because instruments can be built. That is what Erised is for: an understanding of the world at its smallest scales is more than a technical milestone; it is a means to help the world, and the living things in it.
Erised AI builds AI-native systems and world models for nanoscience and metrology: instruments that get smarter with every measurement, because the measurement and the model are designed together. I co-founded it and am its CTO, leading the AI, data and computational microscopy side.
The argument above is the one in my thesis introduction; the note makes it at length.
Models whose parameters can be read as measurements: one learned axis, one meaning, so a fit says something about the system and not only about the prediction.
Semantic-cortical components, NeurIPS 2026 oral · Sparse concept atoms, NeurIPS 2025 · Birdsong in state space
One rich recording in place of many narrow experiments: how cortex represents number, space and time, mapped from people listening to stories.
A model inside the instrument: a brain implant that decodes movement as it records and adjusts Parkinson’s therapy in real time, in people’s homes rather than a clinic.
Nature Biomedical Engineering 2026 · Parkinson’s, measured at home
If you are building, buying or backing nanoscale measurement, or want to work on the models behind it, I would like to hear from you.
Google Scholar LinkedIn GitHub Erised AI
Outside work: philosophy, ballet, hiking, and hosting friends for dinner.