I'm an engineer at heart, drawn to understanding how intelligent systems learn by rebuilding them from first principles. I build small, reproducible experiments in generative AI, optimization, and efficient learning, and write up what works, what fails, and what might come next.
The aim is simple: to really understand how neural networks learn, and how they talk to each other.
forward passgradienthover a neuron · move the cursor to feed the inputs
This network is real and it is trained on this blog. Every deploy, a script reads my posts, cuts them into passages, and trains a 8 → 10 → 10 → 3 network with hand-written backprop and Adam to work out which post a passage came from. The weights drawn above are the trained weights and the output neurons are my posts, so it changes every time I publish.3 posts · 685 words · 127 vocabulary · 162 passages · trained 100% / held-out 68% (chance 33%) · 2026-07-28
My code lives on GitHub. New posts show up in the RSS feed.