NYU Center for Data Science · Machine learning for neural data and biological systems
A severe illness during college changed my life and sent me searching for explanations. That search led me to neural systems, computational neuroscience, and machine learning. I want to use the strongest technical tools I can build to understand difficult biological problems and help make suffering more bearable.
- Personalized handwriting decoding from wrist EMG — Reproduced Meta’s released model and reduced character error from 30.07% to 15.02% across ten held-out users, combining per-user calibration with a language model evaluated after prompt-contamination controls.
- Speech decoding from MEG — 8th place in LibriBrain 2026, with public scores of 0.72230 (Broad) and 0.72929 (Deep). A separate local timing-only baseline reached 0.632; it is not directly comparable to the competition scores.
- Whole-brain perturbation responses in C. elegans — Compared neuron-identity and connectome-based models on unseen animals, with evaluation limits and split details documented in the repository.
- What specifies the C. elegans connectome? — Tested contact geometry, developmental timing, and molecular features against matched controls; broad hidden-code hypotheses did not survive those controls.
- Clinical trial outcome modeling — Built time-aware models and audited label leakage, target encoding, and generalization across indications.
- Physics-based humanoid motion — Used MuJoCo and optimization to derive jumps, throws, and other motions from robot geometry and motor limits, without human motion demonstrations.
I try to make the evidence inspectable: each project links its code, data choices, evaluation, and limitations.
More: About me · Projects · Browser tools · ORCID