4 lessons
Overfitting, underfitting, and regularization
Why a model that's perfect on training data can be useless in the real world, and the math that fixes it.
Evaluation metrics, and why accuracy lies to you
Precision, recall, F1, ROC/AUC, and the confusion matrix, motivated by a classifier that looks great and is actually useless.
The bias-variance trade-off
The single framework that explains why models underfit or overfit, and how it ties overfitting, regularization, and model complexity into one idea.
Data preparation and hyperparameter tuning
What happens before training even starts - handling missing values, encoding categories - and how to systematically search for the best settings.