Longhorn Neurotech

A student organization building EEG and EMG models small enough to run on a prosthetic's own hardware. I've been a member since my first year:

First year, ML Interpretability team: our best EEG classifier reached 85.71% on its test set
Now Edge AI Team Lead for a team of 10, split across data processing, the EMG and EEG models, and quantization with ONNX Runtime
We tested several models on a Raspberry Pi Zero 2 W, including XGBoost and a CNN-LSTM; a small MLP performed best
After INT8 quantization, that EMG gesture model runs in 0.43 ms per sample instead of 1.04 ms and fits in 311 KB, with accuracy going from 90.50% to 90.37%

Edge AI Team Lead, Sep 2025 – presentAI/ML Developer, Sep 2024 – Aug 2025

Spec

EMG input: 8-channel, hand gestures
EMG model: MLP, static INT8 + ONNX
Latency: 0.43 ms/sample vs 1.04 ms FP32
Hardware: Raspberry Pi Zero 2 W
Model size: 311 KB, from ~1.16 MB FP32
Accuracy: 90.50% → 90.37%
Test set: 4,497 windows
EEG work: Capsule network, BCI IV-2a

Fig. 1Our EMG gesture MLP before and after static INT8 quantization with ONNX Runtime, measured on a Raspberry Pi Zero 2 W over 4,497 test windows. Latency drops 2.4× and the model shrinks to 311 KB, for a 0.13-point accuracy cost.

Why I joined

I became interested in brain-computer interfaces after reading about advances like Neuralink's. In my first year I joined the ML Interpretability team, where I built capsule networks for EEG and reordered the channels to match electrode positions on the scalp. That work led to my promotion to Edge AI Team Lead.

The hard part

The hardest part was leading ten people who all needed the same hardware. Everyone wanted to test their own models on the Raspberry Pi, so I set up Raspberry Pi Connect, which let anyone run inference on it remotely from their own laptop. That let us compare models without sharing a single physical board, and it kept everyone involved. On the technical side, ONNX Runtime was the hurdle: PyTorch's built-in quantization made the model larger than the original, so I switched to ONNX Runtime's static quantization, which brought it down to 311 KB.

EMG edge inference on GitHub EEG interpretability on GitHub

Next projectEmbedded Multiplayer Racer →