machine learning

Picking up from hash-based splitting, this post tackles the another hurdle: ensuring your train/test buckets actually represent your target population. Learn how to strategically use stratified sampling, why domain knowledge is irreplaceable, and how to safely combine stratification with SMOTE without poisoning your evaluation.

Why relying on random_state for reproducible train/test splits breaks the moment your dataset changes — and how hash-based splitting with MurmurHash3 gives you deterministic, stable assignments that survive new data.

In this third post of the Learning Machines series, let's examine online versus batch learning, instance-based versus model-based generalization, and write our first linear regression model in scikit-learn.

For years, I wrote code by telling the computer exactly what to do. Shifting to machine learning means letting the data write the rules. This post explores what 'learning' actually means in code, using Tom Mitchell's equation and the supervision taxonomy.

I've been using AI without truly understanding it — calling APIs, fine-tuning models, shipping features powered by things I couldn't fully explain. That gap has started to bother me. This is the first post in Learning Machines, a series where I rebuild my ML and AI foundations from scratch and share the process as I go.