You can build with language models for a long time before the gaps start to hurt. Then a stakeholder asks why accuracy is ninety-four percent but the feature is useless, or a model degrades quietly over six months, and suddenly you need the foundation you skipped. This course is that foundation, written for developers rather than for future researchers.
It is maths-light and code-heavy. You will build a linear model from scratch in NumPy to see what fitting actually means, then move to scikit-learn and stay there. Supervised learning, the train and test split and why it exists, cross-validation, overfitting and underfitting seen in real curves rather than described in the abstract, regularisation, and the bias and variance trade explained through experiments you run yourself.
Metrics get an unusually long treatment because that is where developers are most often embarrassed. Precision against recall on a genuinely imbalanced dataset, the confusion matrix as a decision tool, thresholds and what moving one costs, ROC and precision-recall curves, calibration, and the reason accuracy is close to meaningless when one class dominates.
The rest covers feature engineering, leakage and how to spot it, decision trees and gradient boosting which still win most tabular problems, a brief tour of neural networks, and deployment: monitoring, drift detection and retraining triggers.