cs @ cmu · computational neuroscience & applied ml
i build models of how people respond to things — and spend about as much time checking whether those models measure what they claim to.
| project | description |
|---|---|
| Presonance | comp-neuro model that predicts how an audience responds to a creative before it ships |
| OncoScan AI | oral cancer detection via computer vision — patent pending, incubated, aws-backed |
| medsim | agent swarm over a 3d gaussian-splat world model for trauma center safety — 🏆 most innovative, 2026 harvard hsil hackathon |
| context-ide | dependency-free terminal workspace for running several ai coding agents in one continuous context |
| armasai | assistive-limb behavior in simulation: prompt → morphology → mujoco → policy → metrics |
| paper | what it shows |
|---|---|
| alignment-redistribution | two published papers disagree about whether instruction-tuning makes an llm more brain-like. neither can be right: each model's alignment map is reliable (0.91–0.95), their difference is not (0.005–0.031). you'd need ~490–3,400 h of scanning per subject to settle it. ships a one-correlation diagnostic. |
| braidyn-decoding | a neural decoder built from other animals ages slower than one built from your own earlier data (+0.017 auc, p=5e-5). replicates on an independent two-photon cohort from another lab, survives a cnn and a gru, and isn't a data-volume effect. |
- ml research @ institute for systems biology (baliga lab) — hiv & preeclampsia diagnostic models
- 1st place wa state science & engineering fair — 3d cnn brain stroke classifier
- director of expansion @ ai valley — events with openai, google, aws, nvidia, amd
- funded trader @ topstep ($150k)
python · pytorch · numpy/scipy · scikit-learn · typescript · c++ · r · fastapi · mujoco · aws · gcp