How to interview machine learning engineers
A practical interview loop for ML and LLM engineers: what to test in each round, questions that separate production experience from notebook experience, and red flags.
5 October 2026 · 7 min read
Start from the job, not the title
“ML engineer” covers research-leaning modelling roles, production ML platform roles and LLM application roles. Each needs a different loop. Write down which one you are hiring for and the two or three problems the person will own in their first six months.
If you cannot name those problems, you are not ready to interview. Candidates notice, and the best ones drop out.
Round 1: a production deep-dive (60 minutes)
Ask the candidate to walk through one model or system they shipped. Spend the hour going deeper, not wider: how was the data collected, how was it evaluated, what happened after launch, what broke, and what would they do differently.
Strong candidates talk about evaluation and failure without being prompted. Weaker candidates stay at the level of architectures and libraries.
Round 2: a practical exercise (60–90 minutes)
Use a realistic, bounded problem: design an evaluation set for a retrieval system, debug a training run with a data leak, or review a pull request that adds a feature to a model service. Avoid multi-day take-homes; they filter out people with jobs and families, not weak engineers.
Round 3: hiring manager and team fit (45 minutes)
Focus on ownership and communication. Ask about a time they disagreed with a product or research decision and what they did. Ask how they explain model behaviour to non-ML stakeholders.
Red flags
Cannot explain how they knew a model was working in production. Talks only about accuracy, never about the cost of errors. Has never looked at the raw data. Treats an LLM as the answer to every problem.