Years after our college capstone, an out-of-the-blue email from European AI researchers wanting to reproduce our sign language paper sent me to Google Scholar, where we've been cited 11 times.
Writing
Notes from the workbench
Mostly about security, low-level systems, machine learning that has to survive contact with reality, and the constraints that made a decision obvious in hindsight.
Subscribe via RSSLooking back at my senior capstone project at Missouri State — building a real-time American Sign Language smart home controller with MediaPipe, custom CNNs, and networking — and how it sparked my path in machine learning.
The neural-cost v0.2.0 GPU benchmark moves from CPU roofline analysis to Apple Silicon's MPS GPU backend — measuring PyTorch and JAX across five architectures, four batch sizes, and both eager and compiled execution modes.
An empirical investigation comparing PyTorch 2.14, JAX 0.11, and TensorFlow across five model architectures and four batch sizes, measuring compiler speedups, kernel overheads, and hardware limits on Apple Silicon.
A framework-neutral Python library that estimates a neural network's FLOPs and tensor traffic, measures its runtime, and uses a roofline model to tell you exactly where performance is being left on the table.
The old site was a single 41,000-character index.html. Every change was archaeology. Here is what I replaced it with and why the constraint was maintenance, not design.
Let's connect
Want to talk about any of this?
Most of these posts started as a conversation. Happy to have the next one with you.
Usually replies within a day or two · St. Louis, Missouri