Explore it live
IML Visualizer
Play around and watch what happens when a model trains — a 3D machine-learning lab, built in a single morning.
I was studying gradient descent for my machine-learning course and got tired of imagining loss landscapes. So one morning I built a place to see them.
The visualizer is a full interactive 3D lab: pick from six optimizers, five loss functions, and four models — including a neural network whose backward pass I wrote by hand — and watch the optimization walk across a three.js loss landscape in real time. Rewind recorded runs. Slice the surface. Trace contours computed by marching squares. Every concept has an in-context explanation.
And it’s correct: every hand-written gradient is verified against finite differences to around 1e-10. If the ball rolls somewhere, that’s genuinely where the math goes.
About 14,300 lines, built in one morning, deployed as a static build — which is why it can run live on this page. It’s the clearest example of my favorite loop: when I study something, I build the tool I wish existed, and then the studying takes care of itself.
Want someone who builds like this?
I'm open to opportunities. Tell me what you're building — I answer every email.