Blog

Short explainers on how the boring-but-important parts of machine learning actually work. First-principles, diagrams, pseudocode you can port.

Adversarial Validation

A ten-line trick for spotting data leaks and distribution shift. Train a classifier to tell train rows from test rows, and whichever features it leans on are where your problem lives.

Feature Importance and Partial Dependence Plots

Cracking open the black box. Which inputs your model is actually using, how each one bends the prediction, and why the lazy scatter plot lies. Interactive widget, scribble diagrams.

Weighted Ensemble L2

What AutoGluon stacks on top of everything else. Not L2 regularization — layer 2. Not sklearn — pure NumPy. A twenty-year-old greedy trick that fits in thirty lines and almost always wins.