# 30 machine learning engineer interview questions, and what a good answer shows

> Thirty machine learning engineer interview questions on fundamentals, features, models in production and LLM applications, each with what a strong answer covers.

Source: https://hintvo.com/blog/machine-learning-engineer-interview-questions | Last updated: November 20, 2026 | Reviewed: 2026-11-20

Machine learning engineering interviews test whether your models hold up outside a notebook: honest evaluation, features that are the same in training and production, and systems you can monitor and roll back. Below are 30 common questions, grouped by topic, each with what a good answer shows.

## Fundamentals

- **Explain the bias and variance trade-off with an example.** A good answer describes underfitting and overfitting, what training and validation error look like in each case, and which changes help.
- **What is data leakage, and how have you caught it?** A good answer explains information from the future or from the label slipping into the features, splits by time or by entity, and treats a result that looks too good as a warning.
- **How do you choose an evaluation metric?** A good answer starts from what each kind of mistake costs, explains precision against recall, why accuracy misleads on imbalanced data, and when calibrated probabilities matter.
- **How do you handle an imbalanced dataset?** A good answer picks the right metric first, then class weights, resampling or a tuned threshold, and keeps the real class balance in the validation data.
- **Why keep a validation set and a separate test set?** A good answer tunes on validation, reports once on the test set, uses cross-validation for small data and time-based splits for data that changes over time.
- **When would you pick a simple model over a complex one?** A good answer picks it when the quality is close and it is easier to explain, serve and debug, and always starts from a baseline.
- **What is regularization, and how does it help?** A good answer explains penalizing complexity, with L1, L2, dropout or early stopping, to generalize better, and that L1 can remove features entirely.

## Data and features

- **How do you handle missing values?** A good answer first asks why values are missing, then imputes, adds a missing flag or uses a model that handles them, and never fits the imputation on test data.
- **What is training and serving skew, and how do you prevent it?** A good answer explains features computed differently in training and in production, shares one feature implementation, and logs the features actually served.
- **How do you build features from event data without leaking the future?** A good answer uses point-in-time joins, so each training example only sees data from before its own timestamp.
- **How would you get labels for a new problem?** A good answer writes clear labeling guidelines, runs a pilot to measure how often labelers agree, checks quality and labels the most useful examples first.
- **How much data do you need?** A good answer builds a learning curve from a pilot, and checks whether more data or better labels actually move the metric.

## Models in production

- **Batch or real-time inference?** A good answer weighs the latency the product needs, how fresh the features must be and the cost, and precomputes scores in batch unless the decision depends on the moment.
- **How do you monitor a model in production?** A good answer watches input drift, the distribution of predictions, latency and errors, and the real metric once labels arrive, with alerts that have an owner.
- **What is model drift, and what do you do about it?** A good answer explains that the data or the relationship it models changes over time, detects it, retrains on a schedule or a trigger, and compares the new model with the current one before switching.
- **How do you roll out a new model safely?** A good answer runs it in shadow mode first, then a canary or an A/B test on a business metric, with a quick rollback.
- **How do you make training reproducible?** A good answer versions data and code, fixes random seeds where it can, tracks hyperparameters and the environment, and registers every model it ships.
- **How would you reduce inference latency?** A good answer considers a smaller or distilled model, quantization, batching, caching frequent requests and the right hardware, measured one at a time.
- **How do you decide whether to retrain?** A good answer measures the drop on a fresh labeled sample and weighs it against the cost of retraining and validating.

## LLM applications

- **How would you evaluate a feature built on a large language model?** A good answer builds a test set of real tasks with clear pass criteria, combines automated checks with human review, and tracks regressions on every change.
- **When would you use retrieval instead of fine-tuning?** A good answer uses retrieval for knowledge that changes or must be cited, fine-tuning for format, tone or a narrow task, and says that many systems use both.
- **How do you reduce made-up answers?** A good answer grounds answers in retrieved sources, asks for citations and checks them, lets the model say it does not know, and measures the rate. See [why answers should cite their source](https://hintvo.com/blog/why-ai-answers-need-file-and-line).
- **How do you keep the cost of an LLM feature under control?** A good answer uses smaller models for easy steps, caches, shortens prompts, sets limits per user and measures cost per task.
- **What is prompt injection, and how do you defend against it?** A good answer explains instructions hidden in input data, treats retrieved text as data, limits which tools the model can call, and checks outputs before acting on them.
- **How do you choose the chunk size for retrieval?** A good answer tests on real questions, uses smaller chunks for precise facts and larger ones for context, and adds overlap and metadata where they help.

## Behavioral, for ML roles

- **Tell me about a model that did worse in production than offline.** A good answer explains why, such as skew, leakage, drift or the wrong metric, how they found out and what they changed.
- **Tell me about a project you stopped.** A good answer shows the evidence, how they told people and what the team learned.
- **How do you explain a model's limits to stakeholders?** A good answer uses plain words, example failures and error rates in business terms, and says what a person should still check. See [explaining technical work](https://hintvo.com/blog/explain-technical-work-to-non-technical-stakeholders).
- **A team wants AI in the product without a clear problem. What do you do?** A good answer starts from the user's problem, builds a baseline without machine learning and runs a small experiment with a success metric.
- **Which technique changed how you work?** A good answer names something specific, how they applied it and the result.

## How to practice these

1. Pick the five questions you would find hardest, not the ones you already know.
2. Answer each one out loud in about two minutes, and record yourself if you can.
3. Compare your answer with the note: did it cover what a good answer shows?
4. For every question, write down one real example from your own work that you could use.

For interviewers: the notes work as a scoring guide. Decide before the interview what a strong answer covers, and score the evidence you heard, not the confidence it was said with. See [how to run a fair technical interview](https://hintvo.com/blog/run-a-fair-technical-interview).

More roles and the questions every engineer gets: [interview questions by role](https://hintvo.com/blog/software-engineer-interview-questions-by-role).

> **Where Hintvo fits:** a practice interview asks questions out loud, gives you a score out of 5 and a stronger version of each answer, and on coding questions you can ask for a hint, the approach or the full solution. Practice interviews need a pack of hours. Use Hintvo only where AI help is allowed. Check the rules of your company, school or interviewer first, and respect a no.

## Related

- [Why an AI answer about your code should name its file and line](https://hintvo.com/blog/why-ai-answers-need-file-and-line.md): How can I trust an AI answer about my own codebase?
- [System design interviews: a structure that holds when you are nervous](https://hintvo.com/blog/system-design-interview-structure.md): How do I structure a system design interview?
- [200 software engineer interview questions by role, and what good answers show](https://hintvo.com/blog/software-engineer-interview-questions-by-role.md): What interview questions do software engineers get for each role, and what makes a good answer?
