ML foundations
The vocabulary and judgement layer. These questions come up regardless of whether the role is classical ML or LLM-focused, because they’re about how you think rather than what you’ve memorised.
| # | File | The question it answers |
|---|---|---|
| 01 | Types of ML problem | turning “we want AI for X” into a problem with a metric |
| 02 | Train / validation / test splits | why the split is where production failures are born |
| 03 | Bias-variance trade-off | “my model isn’t good enough” — diagnosed, not guessed |
| 04 | Overfitting and regularisation | L1 vs L2, dropout, early stopping, and what each actually does |
| 05 | The ML lifecycle | end to end, including the unglamorous parts that decide success |
| 06 | When not to use ML | the answer that signals seniority fastest |
| 07 | Hyperparameter tuning | search strategies, Optuna, and why the score’s source matters more |
| 08 | Interpretability | SHAP, permutation importance, counterfactuals, glass-box models |
| 09 | Fairness and bias | proxies, the impossibility result, where bias comes from |
File 03 is the highest-leverage one: the train-vs-validation-error table answers most “what do I do next” questions without guesswork.