Interview prep
ML Engineer: 7-day sprint
Seven days on what ML engineer loops test - modeling fundamentals, tree ensembles, deep learning, ML system design, and production.
Rapid-fire this sprint0/31 done
Reading mode for lessons:
The job. Turn a business problem into a model that works in production, and keep it working: data, features, training, evaluation, serving, and monitoring.
What the loop usually tests
| Round | What they probe | Where it's covered |
|---|---|---|
| ML fundamentals | Bias-variance, leakage, metrics, regularization, picking a model | Days 1 and 2 |
| Deep learning | Backprop, optimizers, training stability, embeddings, transformers | Days 3 and 4 |
| ML system design | End-to-end design: metrics, data, features, model, serving, monitoring | Day 5 |
| Production / MLOps | Latency, cost, drift, reliability, rollouts | Day 6 |
| ML coding | Implement a model or metric from scratch | Each lesson's implementation challenge |
Day 1: How models fail
- Read: Workflow, splits, and leakage, Overfitting and bias-variance, Classification metrics
- Also read: Imbalanced data
- Labs: underfitting vs overfitting, classification threshold
- Cheat sheets: classification metrics, failure modes
- Be ready to: explain leakage with a real example, choose a metric for an imbalanced problem, and pick a threshold from costs.
Day 2: Models that win on tabular data
- Read: Linear regression, Regularization, Logistic regression, Decision trees, Random forests and boosting
- Also read: Feature engineering, Hyperparameter tuning
- Labs: linear regression, decision-tree builder, gradient boosting
- Cheat sheets: model selection, regularization
- Be ready to: compare bagging and boosting, say why GBDTs beat neural nets on most tabular data, and derive the MSE gradient.
Day 3: Training neural networks
- Read: Forward pass, Gradient descent, Backpropagation, Training and regularization
- Also read: Intro to PyTorch, Debugging neural networks
- Labs: forward pass, gradient descent, backpropagation
- Cheat sheets: optimizers, loss functions
- Be ready to: debug a loss that goes to NaN, explain Adam versus SGD, and say what batch norm and dropout do.
Day 4: Embeddings and transformers
- Read: Word embeddings, Self-attention, Transformer architecture, Vector search
- Also read: BERT
- Labs: self-attention, approximate nearest neighbors
- Cheat sheets: transformer architecture, vector search
- Be ready to: walk through Q, K, V with shapes, explain why attention is scaled by , and trade recall against latency in an ANN index.
Day 5: ML system design
- Read: ML system design, Experimentation and A/B testing
- Also read: Time series
- Labs: recommendation funnel, A/B test power and peeking
- Cheat sheet: ML system design
- Be ready to: run the design framework end to end in 40 minutes on a feed, fraud, or search problem.
Day 6: Production
- Read: Production architecture, Reliability, cost, and observability, Serving LLMs
- Also read: Monitoring and drift, Distributed training
- Labs: production architecture, static vs continuous batching
- Cheat sheet: production reliability
- Be ready to: detect and respond to drift, design a safe rollout (shadow, canary, A/B), and size the serving cost.
Day 7: Mock interview
- Rapid-fire: 20 mixed cards from this sprint, 60 seconds each (open the drill). Reread any lesson behind a shaky card.
- Design, 40 minutes each, out loud: Design the "people you may know" recommender for a social network. Then: Design real-time payment fraud detection with a 50 ms budget.
- Coding, 20 minutes each: logistic regression trained with gradient descent in NumPy; ROC AUC from scores and labels without a library.
- Project story: pick one project from the project ladder and practise its five-minute version: problem, baseline, what you tried, the result in numbers, and what you'd do next.