25 Machine Learning Engineer Domande di colloquio

Walk me through how you'd take a model from a Jupyter notebook to a production API.

How do you detect and handle model drift in production?

Describe your approach to feature engineering for a tabular dataset.

What's the difference between L1 and L2 regularization — when would you use each?

How do you decide which evaluation metric to optimize for a classification problem?

Describe a model you deployed that failed in an unexpected way. What happened?

How do you handle class imbalance in a training dataset?

What's your approach to hyperparameter tuning — grid search, random search, or Bayesian?

How would you design an A/B test to evaluate a new recommendation model?

Describe your experience with LLMs — fine-tuning, RAG, or prompt engineering.

How do you ensure reproducibility in your ML experiments?

What's your strategy for versioning models, datasets, and experiments?

How do you explain a complex model's decisions to a non-technical stakeholder?

Describe your approach to data validation before training a new model.

How would you build a real-time inference API that handles 10,000 requests per second?

What's your experience with vector databases and embedding-based search?

How do you approach model monitoring — what signals do you watch?

Describe a time you reduced model training time significantly. What did you do?

How do you handle sensitive data in ML pipelines — privacy-preserving techniques?

What's your experience with distributed training across multiple GPUs or nodes?

How would you implement a feedback loop to continuously improve a deployed model?

Describe your experience with ML frameworks — PyTorch, TensorFlow, JAX?

How do you decide when a rule-based system is better than a machine learning model?

What's your approach to shadow deployments for safely releasing new models?

How do you keep up with the pace of research in ML?