Machine Learning Interview Prep and Questions 2026

You know that weird feeling when you can train a decent model, explain cross-validation, and build a small project, but the second someone says “ML interview” your brain opens 37 tabs and crashes? Yep. Machine learning interviews in 2026 are not just “what is overfitting?” anymore. They mix coding, statistics, system design, product thinking, MLOps, and the very fun moment where someone asks you to debug a model that is “performing well offline but failing in production.”

If you are applying to places like Google, Meta, Amazon, Microsoft, Apple, NVIDIA, Spotify, Uber, Revolut, Booking.com, or Databricks, you need a prep plan that does not waste your evenings. This guide gives you the questions, topics, salary context, and a practical study plan so you can walk in without sounding like you memorized a glossary five minutes ago.

Why machine learning interviews feel harder in 2026#

ML roles have split into several flavors, and companies are more picky about what they need.

A “Machine Learning Engineer” at Amazon might be closer to backend engineering with model deployment. A “Research Scientist” at DeepMind or OpenAI might be math-heavy and paper-heavy. A “Data Scientist, ML” at Spotify or Airbnb might focus more on experimentation, product metrics, and model interpretation.

Here is what makes 2026 interviews different:

1. AI tools raised the bar

   • Everyone can generate a baseline model now.
   • Interviewers want to see if you understand tradeoffs, failure modes, and production issues.

2. MLOps is no longer optional

   • You may get questions on model monitoring, drift, feature stores, CI/CD, and rollback plans.
   • Even junior ML engineers are expected to know the basics.

3. LLM experience helps, but fundamentals still win

   • Prompting is useful, but companies still ask about bias-variance, gradients, evaluation metrics, and data leakage.
   • If you say “I fine-tuned a model,” expect follow-up questions.

4. System design includes ML-specific messiness

   • You might design a recommendation system, fraud detection service, ranking model, or semantic search pipeline.
   • The interviewer wants to hear data flow, latency, metrics, retraining, monitoring, and abuse cases.

ML job titles and salary ranges in 2026#

Before you prep, know what role you are actually targeting. The interview loop changes a lot by title.

Common US salary ranges

These are realistic 2026 base salary ranges for many mid-level roles. Big tech and AI labs can go higher with equity and bonuses.

• Machine Learning Engineer

  • US: $140k to $230k base
  • Senior at Google, Meta, or Netflix: $220k to $320k base, plus equity

• Applied Scientist

  • US: $150k to $240k base
  • Amazon, Microsoft, and Uber often use this title

• Data Scientist, Machine Learning

  • US: $120k to $200k base
  • Product-focused roles at Airbnb, Spotify, DoorDash, and LinkedIn vary widely

• Research Scientist

  • US: $170k to $300k base
  • OpenAI, Anthropic, Google DeepMind, NVIDIA, and Meta AI can exceed this with equity or bonus

• MLOps Engineer

  • US: $130k to $210k base
  • Strong cloud and platform skills can push this higher

Common EU salary ranges

Europe varies a lot by country, and total compensation is usually lower than top US packages.

• Machine Learning Engineer

  • Germany: €70k to €120k
  • Netherlands: €75k to €130k
  • Ireland: €75k to €130k
  • France: €60k to €110k
  • Spain: €50k to €90k
  • UK: £70k to £140k

• Applied Scientist

  • Germany or Netherlands: €80k to €140k
  • UK: £80k to £160k

• Data Scientist, ML

  • EU: €55k to €110k
  • UK: £60k to £130k

• Research Scientist

  • EU: €90k to €180k
  • UK: £90k to £200k
  • Top AI labs in London, Paris, Zurich, or Amsterdam can pay more

If you are seeing a German startup offer €58k for “Senior ML Engineer” with full ownership of the ML platform, model training, deployment, and data pipelines, yes, your eyebrow is allowed to move.

The typical machine learning interview process#

Most ML interview loops follow the same basic shape, with some company-specific weirdness.

1. Recruiter screen

This is usually 20 to 30 minutes.

Expect:

• Role fit
• Salary expectations
• Location or visa questions
• Your strongest ML project
• Why this company

Do not ramble through every project you have ever touched. Pick one strong story and explain it clearly.

2. Technical screen

This may be live coding, ML theory, or project deep dive.

You might get:

• Python coding
• SQL
• Probability and statistics
• Model evaluation
• ML fundamentals
• A short case study

For MLE roles, coding is often serious. Think LeetCode easy to medium, data structures, Python fluency, and clean implementation.

3. ML depth interview

This is where they test if you actually understand models.

Expect questions like:

• How does gradient boosting work?
• Explain regularization.
• When would you choose precision over recall?
• How do you handle imbalanced data?
• Why might validation performance be misleading?
• How would you detect data leakage?

The best answers include tradeoffs. Interviewers like candidates who say, “It depends, here is what I would check first.”

4. ML system design

Usually for mid-level and senior roles.

Examples:

• Design a recommendation system for Netflix.
• Build fraud detection for Stripe.
• Design a search ranking system for Airbnb.
• Create a content moderation classifier for TikTok.
• Build real-time ETA prediction for Uber.

You need to talk about data, features, model choice, APIs, serving, latency, monitoring, retraining, and metrics.

5. Behavioral interview

Do not ignore this one. Plenty of strong technical candidates lose offers here.

Expect:

• Tell me about a failed model.
• Tell me about a time you disagreed with a stakeholder.
• How did you handle unclear requirements?
• Describe a project where you improved a metric.
• Tell me about a time production broke.

Use short stories with numbers. “We improved click-through rate by 4.2%” sounds better than “the model performed better.”

Core machine learning topics you must know#

You do not need to become a walking textbook. But you do need to explain these topics without hiding behind fancy wording.

Supervised learning

Know the basics cold.

You should be able to explain:

• Linear regression
• Logistic regression
• Decision trees
• Random forests
• Gradient boosting, including XGBoost, LightGBM, CatBoost
• Support vector machines
• k-nearest neighbors
• Neural networks

Common questions:

1. What is the difference between regression and classification?
2. Why might logistic regression be a strong baseline?
3. How does a decision tree split data?
4. What makes random forests less prone to overfitting than a single tree?
5. Why does gradient boosting often perform well on tabular data?

A solid answer is simple, accurate, and practical. For example, “I would start with logistic regression because it is fast, interpretable, and gives a strong baseline. If the data has nonlinear relationships, I would compare it with tree-based models like LightGBM.”

Bias and variance

You will almost certainly get this.

Good version:

• High bias means the model is too simple and underfits.
• High variance means the model is too sensitive to training data and overfits.
• You diagnose this by comparing training and validation performance.

Common fixes:

• For high bias:

  • Add features
  • Use a more flexible model
  • Reduce regularization
  • Train longer, if applicable

• For high variance:

  • Add data
  • Use regularization
  • Simplify the model
  • Use cross-validation
  • Apply early stopping
  • Reduce noisy features

Evaluation metrics

This is where a lot of candidates get exposed.

Know when to use:

• Accuracy
• Precision
• Recall
• F1 score
• ROC-AUC
• PR-AUC
• Log loss
• RMSE
• MAE
• MAPE
• NDCG
• MAP
• Hit rate
• Calibration metrics

Examples:

• For fraud detection at Stripe, recall may matter because missing fraud is expensive.
• For spam detection at Gmail, precision may matter because blocking real emails is painful.
• For medical screening, recall can be critical, but false positives still have cost.
• For ad ranking at Meta or Google, offline AUC is not enough. You need online metrics like CTR, conversion rate, revenue, and user experience metrics.

Data leakage

If you remember one thing, remember this: data leakage makes you look good offline and bad in real life.

Common leakage examples:

• Using future information in training
• Random train-test split on time-based data
• Including labels hidden inside features
• Duplicate users across train and test sets
• Preprocessing before splitting data
• Target encoding incorrectly
• Features created after the prediction time

Interview answer structure:

1. Define the prediction point.
2. Check whether each feature exists at that point.
3. Split data by time or entity when needed.
4. Rebuild preprocessing inside the training pipeline.
5. Compare offline and online performance.

Imbalanced datasets

Expect this for fraud, churn, safety, medical, and moderation roles.

Techniques:

• Use better metrics like PR-AUC, recall at fixed precision, or cost-based metrics.
• Change class weights.
• Oversample minority class.
• Undersample majority class.
• Try SMOTE carefully.
• Tune decision thresholds.
• Collect more minority examples.
• Use anomaly detection if labels are scarce.

Do not say “I would just use accuracy.” That is how you create a model that predicts “not fraud” all day and still looks 99% accurate.

Feature engineering

Even in the LLM era, feature engineering still matters.

Examples:

• User activity in last 7 days
• Average basket value
• Time since last purchase
• Device type
• Location features
• Query length
• Historical conversion rate
• Text embeddings
• Image embeddings
• Aggregates by user, item, or session

Important warning: always discuss feature freshness and leakage. A beautiful feature that is not available at prediction time is not a feature, it is a trap.

Deep learning interview topics#

If you mention PyTorch, TensorFlow, transformers, or fine-tuning, expect questions.

Neural network basics

Know:

• Forward pass
• Backpropagation
• Activation functions
• Loss functions
• Optimizers
• Batch size
• Learning rate
• Dropout
• Batch normalization
• Layer normalization
• Early stopping

Common questions:

1. Why do we use nonlinear activations?
2. What happens if the learning rate is too high?
3. What is vanishing gradient?
4. Why use Adam instead of SGD?
5. What is dropout doing?

Simple answer for dropout: “Dropout randomly disables some neurons during training, which forces the network not to rely too heavily on specific paths. It acts like regularization and can reduce overfitting.”

Transformers and LLMs

For 2026, you should understand transformers even if you are not applying to OpenAI.

Know:

• Tokenization
• Embeddings
• Self-attention
• Multi-head attention
• Positional encoding
• Pretraining vs fine-tuning
• Instruction tuning
• RAG
• Vector databases
• Hallucination
• Evaluation
• Latency and cost tradeoffs

Common questions:

1. What is attention?
2. Why are transformers good for language tasks?
3. What is the difference between fine-tuning and RAG?
4. How would you evaluate an LLM application?
5. How would you reduce hallucinations?
6. How would you handle sensitive data in prompts?

A good RAG answer includes retrieval quality, chunking, embeddings, reranking, citation, access control, freshness, and evaluation. Yes, that is a lot. But if you say only “put documents in a vector database,” the interviewer may start typing faster, and not in a good way.

Coding questions for ML interviews#

For MLE roles, you need coding. Not “I can write notebooks” coding. Actual coding.

Python topics to practice

Focus on:

• Lists, dicts, sets, tuples
• Sorting
• String handling
• File parsing
• Classes
• Iterators
• NumPy
• pandas
• Basic algorithms
• Time complexity
• Clean function design

Common tasks:

1. Implement train-test split.
2. Implement precision, recall, and F1.
3. Normalize a matrix with NumPy.
4. Parse logs and compute user sessions.
5. Find top K frequent items.
6. Build a simple k-nearest neighbors classifier.
7. Implement gradient descent for linear regression.
8. Remove duplicates while preserving order.
9. Join two datasets using dictionaries.
10. Compute moving averages over time.

SQL questions

Data roles and product ML roles often include SQL.

Practice:

• Joins
• Window functions
• Group by
• Having
• CTEs
• Date filtering
• Ranking
• Cohort analysis
• Deduplication
• Conversion funnels

Example SQL interview question: “You have an events table with user_id, event_name, timestamp. Calculate the 7-day retention rate for users who signed up in January.”

You should clarify:

• What counts as signup?
• What counts as return?
• Is day 7 exact or within 7 days?
• What timezone?
• Are duplicate events possible?

That clarification habit makes you look senior.

Machine learning system design#

This is the section that scares people, mostly because there is no single perfect answer.

The trick is to follow a structure every time.

ML system design answer framework

Use this order:

1. Clarify the goal

   • What are we optimizing?
   • Who are the users?
   • What is the business metric?
   • What are the constraints?

2. Define prediction target

   • What exactly does the model predict?
   • At what time?
   • How soon is the label available?

3. Data sources

   • User data
   • Item data
   • Event logs
   • Transactions
   • Text, image, audio
   • Third-party data, if allowed

4. Features

   • Real-time features
   • Batch features
   • Historical aggregates
   • Embeddings

5. Modeling approach

   • Baseline
   • Candidate models
   • Training setup
   • Offline metrics

6. Serving

   • Batch or real-time
   • Latency requirements
   • API design
   • Caching
   • Fallbacks

7. Evaluation

   • Offline metrics
   • Online A/B tests
   • Guardrail metrics

8. Monitoring

   • Data drift
   • Prediction drift
   • Label drift
   • Latency
   • Errors
   • Business metric drops

9. Retraining

   • Schedule
   • Trigger-based retraining
   • Human review
   • Rollback plan

Example: Design a recommendation system for Netflix

Here is a strong outline.

• Goal:

  • Recommend movies and shows users are likely to watch and enjoy.
  • Optimize watch time, completion rate, retention, and satisfaction.

• Data:

  • Viewing history
  • Searches
  • Likes and dislikes
  • Watch duration
  • Genre preferences
  • Device type
  • Time of day
  • Region
  • Content metadata

• Candidate generation:

  • Collaborative filtering
  • Similar users
  • Similar items
  • Embedding retrieval

• Ranking:

  • Gradient boosted trees or neural ranking model
  • Inputs include user features, item features, context, and candidate scores

• Metrics:

  • Offline: NDCG, recall@K, MAP
  • Online: watch time, completion rate, return visits, thumbs up, churn

• Serving:

  • Precompute candidates
  • Rank in real time
  • Cache common results
  • Add fallback for new users and new content

• Monitoring:

  • CTR drop
  • Watch time drop
  • Diversity problems
  • Popularity bias
  • Cold start performance

Notice how this answer sounds practical, not academic. That is what you want.

Behavioral questions for ML roles#

Behavioral interviews are not therapy, but they do test whether you are safe to put near a real product.

Use STAR:

• Situation
• Task
• Action
• Result

Keep it short. No one needs an 11-minute documentary.

Common behavioral questions

Prepare stories for:

1. Tell me about a model that failed.
2. Tell me about a time your offline metrics did not match production.
3. Describe a disagreement with a product manager.
4. Tell me about a time you had messy data.
5. Describe a time you had to explain ML to non-technical people.
6. Tell me about a project you led.
7. Tell me about a time you improved latency or cost.
8. Describe a time you found data leakage.
9. Tell me about a time you had to change your approach.
10. Tell me about your most impactful ML project.

What a good story sounds like

Bad: “I built a churn model and it worked well.”

Better: “At my last company, we had a churn model with 0.82 ROC-AUC, but sales did not trust it. I found that many top predictions were customers already in renewal conversations, so the model was not useful operationally. I redefined the target to predict churn risk 60 days before renewal, removed leakage features, and added account activity trends. The AUC dropped to 0.76, but the sales team acted on the list, and retention improved by 3.8% in the pilot.”

That answer is gold because it shows maturity. You understand that useful beats pretty.

The best machine learning interview questions to practice#

Here is your practice bank. Do not just read these. Answer them out loud.

ML fundamentals

1. Explain overfitting and underfitting.
2. What is regularization?
3. L1 vs L2 regularization, when would you use each?
4. What is cross-validation?
5. When is random splitting a bad idea?
6. Explain gradient descent.
7. What is the difference between bagging and boosting?
8. How does random forest work?
9. How does XGBoost work at a high level?
10. What is calibration?

Metrics and evaluation

1. Precision vs recall, explain with an example.
2. When is ROC-AUC misleading?
3. Why use PR-AUC for imbalanced data?
4. How do you choose a classification threshold?
5. What is log loss?
6. RMSE vs MAE, which is more sensitive to outliers?
7. How would you evaluate a ranking model?
8. How would you evaluate a recommendation system?
9. What are guardrail metrics in A/B testing?
10. How do you know if a model improved the product?

Data and features

1. What is data leakage?
2. How do you handle missing values?
3. What do you do with outliers?
4. How do you encode categorical variables?
5. What is target encoding, and what can go wrong?
6. How do you handle high-cardinality features?
7. How do you build time-based features safely?
8. How do you detect training-serving skew?
9. What is feature drift?
10. How would you debug a sudden drop in model performance?

Deep learning and LLMs

1. Explain backpropagation.
2. What is an activation function?
3. Why does batch normalization help?
4. What is attention?
5. What is the difference between encoder and decoder models?
6. Fine-tuning vs RAG, when would you choose each?
7. How do you evaluate generated text?
8. How do you reduce hallucinations?
9. What are embeddings?
10. How would you monitor an LLM app in production?

MLOps

1. How do you deploy a model?
2. Batch inference vs online inference.
3. What is a feature store?
4. How do you version data and models?
5. How do you monitor model drift?
6. What is a shadow deployment?
7. What is a canary release?
8. How would you roll back a bad model?
9. How do you test ML pipelines?
10. How do you manage model retraining?

30-day machine learning interview prep plan#

If you have a month, do this. Not perfect, just consistent.

Week 1: Refresh fundamentals

Daily:

• 45 minutes ML theory
• 45 minutes coding
• 20 minutes flashcards or notes

Topics:

• Bias-variance
• Linear and logistic regression
• Trees and boosting
• Metrics
• Cross-validation
• Leakage
• Imbalanced data

Output:

• One-page notes for each topic
• 10 spoken answers recorded on your phone

Yes, record yourself. You will hate it for 2 minutes, then realize it works.

Week 2: Coding and SQL

Daily:

• 1 coding problem
• 1 SQL problem
• 30 minutes NumPy or pandas practice

Focus:

• Implement metrics
• Parse data
• Aggregations
• Window functions
• Clean Python functions
• Complexity explanation

Output:

• 15 solved Python problems
• 10 solved SQL problems
• 3 mini implementations from scratch, like logistic regression, KNN, or gradient descent

Week 3: System design and projects

Daily:

• 1 ML system design prompt
• 30 minutes project review
• 20 minutes behavioral story practice

Prompts:

• Recommendation system for Spotify
• Fraud detection for Revolut
• Search ranking for Airbnb
• ETA prediction for Uber
• Churn prediction for Salesforce
• Content moderation for TikTok
• Demand forecasting for Amazon

Output:

• 5 structured design answers
• 3 project deep dives
• 5 STAR stories

Week 4: Mock interviews and cleanup

Daily:

• 1 mock interview
• Review weak spots
• Tighten resume bullets
• Practice compensation answer

Do:

• 2 coding mocks
• 2 ML theory mocks
• 2 system design mocks
• 1 behavioral mock

Also prepare your intro: “I’m an ML engineer with X years of experience building models for Y. My strongest work is around Z, where I improved metric A by B%. I’m now looking for roles where I can build production ML systems with strong product impact.”

That is clean. Please do not start with where you went to primary school.

How to explain your ML projects in interviews#

Your project explanation needs structure. Otherwise, you will drift into notebook-tour mode.

Use this:

1. Problem

   • What was the business or user problem?

2. Data

   • What data did you use?
   • How big was it?
   • What were the quality issues?

3. Target

   • What exactly were you predicting?

4. Approach

   • Baseline
   • Models tried
   • Feature work
   • Validation strategy

5. Result

   • Offline metrics
   • Online metrics
   • Business impact

6. Production

   • Deployment
   • Monitoring
   • Retraining
   • Failures

7. What you would improve

   • Better labels
   • More data
   • Simpler model
   • Better monitoring
   • Cost reduction

Example project pitch

“I built a lead scoring model for a B2B SaaS sales team. The goal was to rank inbound leads by conversion probability so reps could prioritize faster. We used CRM data from Salesforce, website events from Segment, company firmographics, and email engagement data. I trained a LightGBM model after starting with logistic regression as a baseline. The key issue was leakage from sales activity after lead creation, so I rebuilt features based only on data available at scoring time. Offline PR-AUC improved from 0.31 to 0.46, and in a four-week pilot, qualified meetings increased by 12% for the same rep capacity.”

That is the kind of answer that gets follow-ups instead of suspicion.

Common mistakes to avoid#

Please avoid these. They are painfully common.

1. Memorizing definitions without examples

If you define precision but cannot explain it using fraud, spam, or medical screening, you are not ready.

2. Ignoring baselines

A simple baseline is your friend. Interviewers like candidates who start simple and improve with evidence.

3. Saying “deep learning” for every problem

For tabular business data, LightGBM might beat your fancy neural net and cost 1% as much.

4. Forgetting latency and cost

A model that takes 4 seconds to return recommendations may be useless for a consumer app. Mention latency like a grown-up.

5. Not asking clarifying questions

If they ask, “Build a churn model,” do not jump into algorithms.

Ask:

• What is churn?
• What is the prediction window?
• Who uses the prediction?
• What action will they take?
• What is the cost of false positives and false negatives?

6. Pretending you know everything

If you do not know, say how you would reason through it. Interviewers respect honesty more than confident nonsense.

Final checklist before your ML interview#

The night before, review this list.

You should be able to:

• Explain your top 2 projects in 3 minutes each
• Answer bias-variance clearly
• Explain precision, recall, F1, ROC-AUC, and PR-AUC
• Spot data leakage examples
• Discuss imbalanced classification
• Write clean Python for basic data tasks
• Write SQL with joins and window functions
• Design a basic recommendation or fraud system
• Explain batch vs real-time inference
• Talk about monitoring and drift
• Give 5 behavioral stories with numbers
• Ask smart questions about the team and role

Good questions to ask them:

1. How are models deployed here?
2. What does the team own from research to production?
3. What metrics define success for this role?
4. How often do models get retrained?
5. What are the biggest ML reliability issues the team faces?
6. How do data scientists, ML engineers, and product managers work together?
7. What would a successful first 90 days look like?

Quick salary and negotiation tip#

When recruiters ask salary expectations, do not trap yourself too early.

Try: “I’m still learning about the scope of the role, but based on similar ML roles in the market, I’d expect something in the range of $170k to $220k base in the US, depending on level and total compensation. I’m open to discussing the full package.”

For Europe: “Based on similar ML engineering roles in Amsterdam and Berlin, I’d expect something around €85k to €120k base depending on level, equity, and scope.”

Adjust numbers by country, level, and company. OpenAI, Anthropic, NVIDIA, Google DeepMind, Meta, and Netflix can be very different from a 40-person startup with “AI” in the pitch deck and one exhausted data engineer.

You do not need to know everything#

Machine learning interviews are broad, but they are not random. Most questions test the same core things: can you reason with data, can you build models that work outside a notebook, can you communicate tradeoffs, and can you avoid breaking production while looking confident on Zoom.

Prep the fundamentals, practice out loud, build a repeatable system design structure, and get your project stories sharp. You will already be ahead of the person who spent three weeks rereading transformer diagrams and never practiced explaining precision and recall.

Before you apply, make sure your resume is not quietly failing ATS filters. Run it through JobRise’s free checker here: https://jobrise.io/en/free-ats-checker/ and fix the easy stuff before a recruiter ever sees it.