AI & ML Interviews.
Machine learning interview prep — fundamentals, leakage, evaluation, deep learning, LLMs and RAG, and ML system design.
Beginner
Start here — no prior experience needed
The AI and ML Interview
Which of the four ML roles you are actually interviewing for, what each loop tests, and a warm-up question answered two ways.
ML Fundamentals Questions
Bias-variance, regularisation and the bias-variance-free lunch — each demonstrated with measured curves rather than recited definitions.
Feature Engineering and Data Leakage
The four kinds of leakage, each shown as the AUC it inflates — plus the scaling mistake almost everyone makes and the time-series split that hides it.
Model Evaluation Questions
Choosing a metric that matches the decision — precision/recall tradeoffs, why ROC AUC misleads on rare events, calibration, and confidence intervals on a score.
The ML Coding Round
Vectorised numpy, pandas without loops, and the algorithms you are asked to write from scratch — logistic regression, k-means and a metric, each timed against the naive version.
Intermediate
For developers with core concepts down
Classical ML Depth
Trees, bagging and boosting — what each actually does to bias and variance, why gradient boosting still beats deep learning on tabular data, and how to read feature importance without being misled.
Deep Learning Questions
Backpropagation, vanishing gradients, normalisation and optimisers — the mechanisms interviewers ask about, each shown as a measurement in PyTorch.
LLM and RAG Questions
Attention costs, chunking, retrieval evaluated with recall@k, and the RAG failure modes — measured, so 'it hallucinated' becomes a diagnosis rather than a complaint.
Advanced
Production-grade patterns for experienced engineers
ML System Design
The 45-minute round where the model is the smallest part — framing the problem, sizing it, choosing an offline metric that tracks the business one, and serving it inside a latency budget.
MLOps and Production Questions
Drift detected with PSI, reproducibility that actually holds, rollback that works, and the retraining trigger that is not a calendar.