The Interview Process
Stage 1: Recruiter screen (30 min)
Background, motivations, and salary. Expect a question on your experience with specific ML frameworks and your most significant model deployment.
Stage 2: Take-home ML problem (2–4 hrs)
A Jupyter notebook task — EDA, feature engineering, model training, and evaluation. Quality of analysis and reproducibility matter as much as model performance.
Stage 3: Technical review of take-home (45–60 min)
You'll walk through your solution and justify decisions. Prepare to be challenged on your feature choices, evaluation approach, and what you'd do differently.
Stage 4: ML system design (45 min)
Design a training and serving pipeline for a real-world problem. Focus on data pipelines, model versioning, feature stores, and monitoring.
Stage 5: Behavioural (45 min)
STAR-format questions about past model deployments, cross-functional work, and handling technical disagreements.
Technical Questions
Write your own answer first, then compare against the example.
Q1. How do you diagnose and address overfitting in a model you've trained?
Strong answer
Start with diagnostics: plot training vs. validation loss curves. If training loss continues to fall while validation loss plateaus or rises, the model is overfitting. Remedies in rough order of effort: (1) Regularisation — L1/L2, dropout for neural networks. (2) Reduce model complexity — fewer layers, smaller embedding dimensions. (3) Add training data — real data if possible, augmentation otherwise. (4) Early stopping based on validation loss. (5) Ensemble methods which average over variance. Also check for data leakage — it can masquerade as overfitting or artificially inflate training performance.
Q2. Walk me through how you would choose between a tree-based model and a neural network for a tabular regression task.
Strong answer
For most tabular regression problems, a gradient-boosted tree (XGBoost, LightGBM, CatBoost) is the right starting point: faster to train, more interpretable, handles missing values gracefully, and consistently wins on tabular benchmarks. Neural networks win when: there's large data volume (millions of rows), features include embeddings or text, or you need to share representations across tasks. I always start with a strong GBDT baseline before investing in neural architectures. The key trade-off is explainability: stakeholders in regulated industries often require feature importances that GBDTs provide naturally.
Q3. Explain precision, recall, and F1 score. When would you optimise for each?
Strong answer
Precision = TP / (TP + FP): of everything the model predicted positive, how many were actually positive. Recall = TP / (TP + FN): of all actual positives, how many did the model catch. F1 is their harmonic mean. Optimise for precision when false positives are costly (e.g. fraud alerts that trigger manual review — too many false alarms waste analyst time). Optimise for recall when false negatives are costly (e.g. cancer detection — missing a true positive is unacceptable). In practice, plot the precision-recall curve and choose the operating threshold based on the business cost of each error type. For imbalanced classes, F1 or area under the PR curve is more informative than accuracy.
Q4. How do you handle class imbalance in a classification problem?
Strong answer
Several strategies, often combined: (1) Resampling — oversample the minority class (SMOTE) or undersample the majority. (2) Class weights — pass class_weight='balanced' to sklearn models so the loss penalises minority class errors more. (3) Use appropriate metrics — accuracy is misleading; use precision-recall AUC or F1. (4) Threshold tuning — the default 0.5 threshold is rarely optimal; tune it on the validation set. (5) Algorithm choice — tree-based methods with class weights often work well out of the box. Always start with reweighting before more complex interventions.
Q5. What is data leakage and how do you prevent it?
Strong answer
Data leakage occurs when information from outside the training time period (or conceptually unavailable at prediction time) is included in features. Common forms: target leakage (a feature that is computed using the target), temporal leakage (using future data in a time-series split), and train-test contamination (fitting scalers/encoders on the full dataset before splitting). Prevention: always fit preprocessing transformers inside cross-validation folds, use time-based splits for temporal data, audit feature definitions against the point-in-time when the model would be called, and sanity-check suspiciously high evaluation scores.
Q6. Describe how you would set up an A/B test to evaluate a new ML model in production.
Strong answer
Define the hypothesis and primary metric before starting. Randomly split traffic (or users, not sessions) into control (current model) and treatment (new model). Determine sample size upfront using a power analysis — underpowered tests produce inconclusive results. Run the experiment for at least one full business cycle (e.g. one week to capture day-of-week effects). Monitor guardrail metrics (latency, error rate) alongside the primary metric. After the experiment, use a statistical test (t-test, Mann-Whitney) and report confidence intervals, not just p-values. Shadow deployment first if the new model carries risk.
Q7. How do you approach feature engineering for a time-series forecasting problem?
Strong answer
Standard time-series features: lag features (value at t-1, t-7, t-30), rolling statistics (rolling mean, std, min/max over a window), date/time components (hour of day, day of week, month, is_holiday). Domain-specific features are often the most valuable: for retail, promotions and competitor prices; for energy, weather. Watch for target leakage — all features must use data available strictly before the prediction timestamp. For neural approaches (LSTM, Temporal Fusion Transformer), raw time series can be fed directly, but hand-crafted features still often boost performance.
Q8. What is the difference between bagging and boosting?
Strong answer
Bagging (Bootstrap AGGregating) trains multiple models independently on random subsets of the data and averages their predictions. It reduces variance — Random Forest is the canonical example. Boosting trains models sequentially, each one correcting the errors of the previous. It reduces bias. Gradient boosting (XGBoost, LightGBM) is typically more accurate on structured data but more sensitive to overfitting and hyperparameters. Bagging is more parallelisable. In practice: start with gradient boosting for accuracy, use Random Forest if you need faster training or built-in parallelism with less tuning.
Q9. How do you validate a model intended for use in a regulated industry (e.g. finance or healthcare)?
Strong answer
Stricter validation requirements than standard ML: (1) Hold out a completely unseen test set that no one looks at until final evaluation. (2) Document the model card: training data sources, known limitations, performance by demographic subgroup. (3) Fairness audits — test for disparate impact across protected characteristics. (4) Adversarial testing — deliberately craft edge cases to stress-test the model. (5) Full audit trail of experiments, hyperparameters, and dataset versions (MLflow or equivalent). (6) Human-in-the-loop for high-stakes decisions. Know which regulatory frameworks apply (GDPR Article 22, EU AI Act risk tiers).
Q10. What is the curse of dimensionality and how does it affect ML model training?
Strong answer
As the number of features grows, the volume of the feature space grows exponentially, meaning data points become increasingly sparse. Effects: distance-based algorithms (KNN, SVM with RBF kernel) degrade because all pairwise distances converge. Models that rely on density estimation become unreliable. Training time grows. Solutions: dimensionality reduction (PCA, UMAP, t-SNE), feature selection (remove low-variance or low-importance features), regularisation to penalise model complexity, and domain expertise to select meaningful features rather than exhaustively adding all available ones.
Behavioural Questions
Use STAR format and keep answers to 2–3 minutes.
Tell me about a model you built that performed well in development but poorly in production. What happened and how did you resolve it?
Shows awareness of distribution shift, monitoring, and the gap between offline metrics and real-world value. Avoid a story where the problem was never solved.
How do you communicate model uncertainty or limitations to a non-technical audience?
ML engineers regularly need to explain what a model can and can't do. Demonstrate empathy for the audience and use concrete examples rather than abstract statistics.
Describe a time you had to decide between model accuracy and business constraints (speed, cost, explainability). How did you make the call?
Shows engineering judgment. The 'right' answer depends on context — the interviewer is testing your reasoning process, not looking for a specific outcome.
How do you keep a machine learning project on track when requirements change mid-sprint?
ML projects are inherently iterative. Describe how you scope experiments, timebox work, and manage stakeholder expectations.
Describe how you've shared technical knowledge with a team that had less ML experience.
Collaboration and knowledge transfer are core skills. Describe a concrete example — documentation, code review, a lunch-and-learn — and what changed as a result.
Red Flags to Watch For
No versioning for datasets or models
Without versioned datasets and model artefacts, experiments are irreproducible and production rollbacks are impossible.
Evaluation on the training set
A basic sign of process immaturity. If the team can't describe their holdout strategy, model quality estimates are meaningless.
No clear owner for model quality
Models degrade over time due to distribution shift. Someone needs to be accountable for monitoring and retraining.
"We'll figure out deployment later"
If engineering and ML haven't agreed on how models reach production, you'll spend months on that problem after training the model.
Optimising the wrong metric
Ask what success looks like for the model. If the answer is accuracy on an imbalanced class dataset, or a metric disconnected from business outcomes, there's a strategic problem.