Long-form notes on what I'm learning, AI evaluation, design engineering, and the craft of shipping.
A resume flattened onto a webpage says 'I can write HTML.' I wanted the medium to be the message, so this site is a workstation you can actually drive.
Three months of grading frontier-model training data changed how I review my own code. The failure patterns are the same, they just hide in different places.
I trained VGG19, MobileNet, ResNet, EfficientNet, and Inception on the same medical-imaging problem. The interesting result wasn't which one won, it was why the ranking flipped when I fixed the data.
DNS tunneling smuggles data out of a network inside ordinary-looking lookups. You can't eyeball it, and there's far too much of it to inspect by hand, so detection has to be statistical, interpretable, and built to scale.