_{My own brain fibers. Traced from my 3T diffusion MRI, in the three canonical views, with a NumPy rasteriser I wrote.}

Biomedical engineering at UT Austin. I build brain-computer interfaces, the neuroimaging pipelines that feed them, and infrastructure for autonomous agents.

Most of my week is EEG and diffusion MRI. The rest is backend work, usually because a pipeline needed something that did not exist yet.

INIaustin.org | reachequitylab.com | moltgrid.net | Interactive card | LinkedIn

Fable decodes meaning instead of motor intent, reading the N400 event-related potential while a story unfolds. Built in 48 hours by a team of five at Global NeuroHack 2026 in San Francisco, where it placed 2nd internationally. There is a live demo.

Longhorn Neural Interface Platform is Longhorn Neurotech's desktop EEG application, covering acquisition, DSP, model training and deployment. 35 released builds so far, currently v1.46, with separate macOS and Windows artifacts and an in-app updater. Supports the g.tec Unicorn Hybrid Black over native BLE, OpenBCI Cyton, Myo, and BrainFlow. I am the top contributor, with 80 of its 121 commits.

ssvep-device-control is a real-time SSVEP interface: 250 ms decision windows, above 90 percent accuracy across 4 degrees of freedom.

Cortex replays recorded primate M1 spikes on Apple Silicon through a CoreML NDT1 decoder, a ReFIT-Kalman filter and a 120 Hz Metal cursor renderer, at 8.8 ms p99 software pipeline latency on an M5 Pro. It reports its own negative results: a ridge decoder on raw spikes beats the 1.3M-parameter transformer on held-out velocity R2 (0.4616 against 0.4238), so the demo runs the ridge decoder by default.

neurorust is a Rust runtime for intracortical data on PREEMPT_RT Linux, built for bounded worst-case latency at implant scale (1024 channels at 30 kHz) rather than good averages. It is early. What exists is the measurement harness every later latency claim rests on: it refuses a run unless fifteen preconditions hold on the live machine, and re-derives each published figure from the raw capture.

The banner is my own diffusion scan. The pipeline behind it takes preprocessed multi-shell data (b = 0 through 3000, 96 directions, 1.7 mm isotropic), reconstructs it with GQI, tracks 600,000 streamlines, and recovers 55 named bundles by atlas recognition. What you see is the fibers themselves, not a whole brain: diffusion MRI tracks water moving along axons, so the grey matter they connect is not in the picture.

The renderers are pure NumPy, no GPU and no 3D library. The banner is drawn in the flattest of them, where every streamline is a hairline of one colour and the tone comes entirely from how much ink lands on each pixel, on a Beer-Lambert curve so crossings darken smoothly instead of clipping. The lit renderer that produced the imagery on INIaustin.org is the same rasteriser with depth-slab compositing, illuminated streamline shading, and a vectorised z-buffer that gets its depth test for free from sorting samples far to near.

INIaustin.org is the site for the UT Austin chapter of the Institute of Neuro Innovation, the second chapter after UCLA. I built the site and the imagery on it.

Alongside that I work on MRI preprocessing infrastructure (dcm2bids through MRIQC, fMRIPrep, QSIPrep, ASLPrep and QSMxT) and first-author exposome research at the REACH Equity Lab, with an abstract accepted to ISES 2026 in the machine learning session.

MoltGrid is a backend-as-a-service for autonomous agents: memory, task queues, inter-agent messaging, scheduling and vector search behind one API. FastAPI, PostgreSQL and Redis, 245 routes, 780 tests, Apache 2.0. It ships typed SDKs for Python and JavaScript, plus an MCP server so agents can call it natively.

donnyclaude is a workflow engine for Claude Code: durable project state, phase planning, and completion gates that fail rather than wave work through. 94 skills, 48 agents, 63 commands and 32 hooks, installed with one command. v3.2.0 on npm, MIT.

• The card, a draggable fsaverage5 cortical mesh that heat-maps when you pick an emotion, using MNI coordinates taken from published meta-analyses rather than invented.
• Infinite Hallway, a procedurally generated corridor that re-centres the whole scene every 2000 units so it never runs out of floating point.
• polaroids-for-mom, a macOS screensaver of polaroids scattered on a wooden table. A Mother's Day gift, packaged so it never makes a network call.
• synq, emotion-scored ad creative that re-optimises against viewer sentiment.