Machine Learning Engineer Interview Questions (2026): 15 Questions

A 5-step way to answer any ML engineer question

Define it. One clear sentence.
Give the intuition. Why it matters in practice.
Show the diagnostic. How you would detect the issue.
Give the fix. Standard remedies and trade-offs.
Connect to production. How it affects deployment or monitoring.

Contents: Fundamentals (questions 1 to 5) · Modeling (questions 6 to 10) · Production (questions 11 to 15)

Fundamentals (questions 1 to 5)

1. What is the bias-variance trade-off?

Bias is error from overly simple assumptions; variance is sensitivity to the training sample.
Simple models underfit (high bias); complex models can overfit (high variance).
Diagnose with training versus validation error.
Fix with more data, regularization or a different model capacity.

Common pitfall: Defining the terms without saying how to diagnose them.

2. Precision, recall and F1: when do you use each?

Precision: of predicted positives, how many are correct.
Recall: of actual positives, how many were found.
F1 is their harmonic mean, useful when both matter.
Choose by error cost: recall for missed fraud, precision for spam filters that should not block real mail.

Common pitfall: Defaulting to accuracy on imbalanced data.

3. What do ROC AUC and PR AUC measure?

ROC AUC measures ranking quality across thresholds using true and false positive rates.
PR AUC focuses on the positive class and is more informative when positives are rare.
Both are threshold-independent; deployment still needs a chosen threshold.
Report the metric that matches the business decision.

Common pitfall: Using ROC AUC alone for highly imbalanced problems.

4. What is data leakage?

Information in training features that would not be available at prediction time.
Examples: future data, target-derived features, duplicates across train and test.
It makes offline results look much better than production.
Prevent with time-based splits and feature audits.

Common pitfall: Normalizing using statistics computed on the full dataset including test data.

5. How do you use cross-validation correctly?

Split data into folds, train on some and validate on the rest, rotating.
Use stratified folds for classification and time-based folds for temporal data.
Keep groups such as users within one fold to avoid leakage.
Fit preprocessing inside each fold.

Common pitfall: Random K-fold on time-series data.

Modeling (questions 6 to 10)

6. Why do gradient-boosted trees often win on tabular data?

They handle mixed feature types, missing values and nonlinear interactions well.
They need less feature scaling and tuning than neural networks.
They train quickly and are strong baselines.
Neural networks shine with unstructured data such as text and images.

Common pitfall: Defaulting to deep learning for small tabular datasets.

7. How do you handle missing data?

Understand why it is missing; missingness can itself be informative.
Impute simply (median, mode) or with models, and add missing indicators.
Some models, such as many tree libraries, handle missing values natively.
Apply the same handling in training and serving.

Common pitfall: Dropping rows without checking how much data or bias that removes.

8. What is regularization?

Adding constraints or penalties to reduce overfitting.
L1 encourages sparsity; L2 shrinks weights smoothly.
Other forms: dropout, early stopping, data augmentation.
Tune strength on validation data.

Common pitfall: Mixing up which of L1 and L2 produces sparse weights.

9. How do you choose features?

Start from domain knowledge and simple signals.
Use importance measures and ablations to keep useful features.
Check for leakage and serving availability.
Remove features that are costly to compute online with little gain.

Common pitfall: Adding features that are not available in real-time at serving.

10. What is model calibration and why does it matter?

A calibrated model's predicted probabilities match observed frequencies.
It matters when probabilities drive decisions, such as pricing or ads auctions.
Check with reliability diagrams or expected calibration error.
Fix with Platt scaling or isotonic regression.

Common pitfall: Treating raw scores as probabilities.

Production (questions 11 to 15)

11. How do you deploy a model?

Package the model with its preprocessing and version it.
Serve via batch jobs or an online API depending on latency needs.
Roll out with shadow mode, canary and A/B tests.
Keep rollback ready.

Common pitfall: Deploying the model without the exact preprocessing used in training.

12. What is data drift versus concept drift?

Data drift: the input distribution changes.
Concept drift: the relationship between inputs and the target changes.
Detect data drift with distribution tests; concept drift needs labels.
Respond with retraining, feature updates or model redesign.

Common pitfall: Treating every distribution change as a reason to retrain immediately.

13. How do you decide when to retrain?

On a schedule, on drift alerts or on quality drops.
Compare the new model against the current one on recent data.
Automate the pipeline but keep evaluation gates.
Track the cost of retraining versus gain.

Common pitfall: Retraining automatically without validation.

14. How do you make ML experiments reproducible?

Version code, data and configuration.
Fix random seeds and record environments.
Track experiments and metrics in a tool.
Store model artifacts with lineage.

Common pitfall: Relying on notebooks with untracked changes.

15. Tell me about a model you put in production. What do interviewers look for?

A clear problem, metric and baseline.
Your specific decisions and trade-offs.
Results measured online, with honest limitations.
What you would do differently.

Common pitfall: Describing the team's work without your own role.

Go deeper on related questions

FAQ

How do ML engineer interviews differ from data scientist interviews?

ML engineer loops weigh coding, systems and production more; data scientist loops weigh statistics, experimentation and analysis.

Do ML engineer roles now include LLM questions?

Many do. Expect some questions on embeddings, RAG or fine-tuning alongside classic ML.

What should I review first?

Metrics, leakage, validation and the bias-variance trade-off, then deployment and monitoring.

How do I practice?

Explain one end-to-end project out loud using the five steps, including the metric you moved.

The PlaybookWant the full method?

The 0 to 1 AI Engineer Interview Playbook covers answer frameworks, system-design walkthroughs and an 8-week plan. $19.99 on Gumroad.

Get the playbook

All interview topics · Prep guides · Question deep dives · Free 50 RAG and LLM questions