Personal Projects

I built a neural network, fed it over two million academic articles, and trained it on my GPU. I hosted the resulting model on Amazon Web Services, so you can give it a seed yourself, then watch it invent new physics, cite yet-to-exist references, and discover and name new chemical elements. The model is hundreds of thousands of times smaller than a modern LLM, but the principles applied here can be scaled up with greater resources.

I built another neural network, this time fed it the bulk of 75,000 images of grains of rice taken at a lab in Türkiye, and trained it on Google's GPUs. Following best practices in training (and having high uniformity in the training data) led to a fairly generalized model — it correctly identifies images it has never seen before with an accuracy of 98% (when those images are of similar uniformity to the training data, of course).

Good trivia games are hard to find because they are hard to design — every player thinks differently — so significant care went into designing the user experience for this. In the absence of an army of testers, it is necessary to place oneself in the shoes of varying personas and approach one's own challenges from creative angles.

The game is fast and fluid and witty — and quite addictive. Although be prepared that this is actually two separate experiences disguised as one.