Machine learning interview questions: what AI companies actually ask

Machine learning interview questions fall into five groups: ML fundamentals, practical coding, ML system design, LLM and evals questions, and a deep dive on your past projects. With a recruiter screen and a values round, that makes four to six stages. Most rounds test whether you can ship and evaluate a model, not solve puzzles.

This guide uses what employers write themselves. On Tokn, as of 14 October 2026, 171 of the 787 AI job postings we matched to their source publish their interview process in the job ad, at 28 of the 79 companies we track. We added official guides from OpenAI, Anthropic, Doctolib, Synthesia and Canva, and what 274 ML and AI engineer ads require. For CVs and applications, see our step-by-step guide to landing an AI job.

In short

• About one AI job ad in five on Tokn spells out its interview process, usually a recruiter screen, a technical deep dive and a values round.
• Coding appears in about half of the technical loops that publish their stages, and several say it is not LeetCode style.
• System design rounds increasingly cover LLM systems: retrieval, evals, cost and latency.
• AI tools: Canva expects them, Synthesia allows them in its coding round, Anthropic bans them in live interviews unless told otherwise. Ask first.

What does an ML engineer interview look like?

A typical ML engineer interview has a recruiter screen, a hiring-manager call, one to three technical rounds, a values or behavioural round and sometimes a reference check. The table counts stages in the 94 technical postings on Tokn that publish their process.

This describes the companies that publish, not the market. Wayve makes up 26 of the 94 postings, hence the last column. Many labs run loops without printing them: none of OpenAI's 50 postings describes one.

Interview loops at European AI companies

Loops as printed in job ads advertised in October 2026:

Mistral AI publishes several loops; we cover them in Mistral AI's interview loops.

How long the process takes

Published targets run from 10 days to three weeks: Qonto aims for "10 days max", Synthesia says "we can be done in two weeks", and a Shift Technology internship ad said "2 to 3 weeks". OpenAI's interview guide, as of October 2026, describes "4–6 hours of final interviews with 4–6 people over 1–2 days".

Machine learning interview questions on fundamentals

Fundamentals rounds test whether you understand why models fail. At least 39% of ML and AI engineer requirement lists on Tokn mention maths or ML fundamentals (71 of 180), and a Mistral AI data scientist loop had a 60-minute "Python + Stats/ML" round.

1. Explain the bias–variance trade-off. How do you spot overfitting? Compare training and validation curves. Overfitting: more data, regularisation, early stopping. Underfitting: a richer model.
2. Your fraud model has 99% accuracy. Is it good? Not if fraud is rare: predicting "no fraud" scores about the same. Use precision, recall, PR-AUC and a cost-based threshold.
3. Give an example of data leakage and how you would catch it. Future information in features, duplicates across splits, target encoding before the split. Use time-based splits; distrust scores that look too good.
4. What causes vanishing or exploding gradients? Deep stacks and saturating activations. Fixes: residual connections, normalisation, careful initialisation, gradient clipping. OpenAI's guide recommends the Deep Learning book for this ground.

Coding questions for ML roles

The coding rounds companies describe are practical Python on data and model code, not puzzles. Synthesia's is "not Leetcode style", H Company's are "modelled on real-world problems", and Wayve tests PyTorch debugging. Python is in the skills of 69% of the 274 ML and AI engineer roles on Tokn, PyTorch in 31% (more in the tools AI job ads ask for).

5. This PyTorch training loop doesn't learn, or the loss goes NaN. Find the bugs. Missing zero_grad, probabilities passed instead of logits, model.eval() left on, a bad learning rate. First check it can overfit one batch.
6. Implement multi-head attention with a causal mask. Shapes, scaling by √d, masking before the softmax, reshaping heads. LLM internals appear in at least 28% of requirement lists.
7. Live SQL or Python on an events table: compute 7-day retention by cohort. Mistral AI's Senior Data Engineer loop had a 45-minute Python and SQL live coding round.
8. Review this pull request, or fix this performance bug. Reproduce, profile, make the smallest safe fix, add a test. Synthesia's hands-on round includes both.

ML system design interview questions

An ML system design interview is a 45 to 75 minute conversation where you design a model-backed product end to end. Synthesia's interview preparation guide, linked from its job ads, says its ML design round "is focused on GenAI and LLM use cases, validation of outputs, cost, latency". It suggests a flow worth reusing:

• Requirements and success metric (5–10 minutes)
• High-level architecture (15 minutes)
• Deep dives and trade-offs (15 minutes)
• Wrap-up, including monitoring (10 minutes)

Its listed mistakes include staying silent and rushing into code; Doctolib's interview guide asks you to "talk through your approach, assumptions, and trade-offs."

9. Design a RAG assistant over internal documents. Chunking, embeddings, hybrid search, reranking, citations, evals for recall and faithfulness, a cost budget. RAG or retrieval appears in at least 44% of requirement lists.
10. Design a system that extracts fields from invoices and suggests accounting entries. OCR plus a layout or vision-language model, confidence thresholds, human review, per-field evals. It mirrors Pennylane's case study "covering problem framing, data, modeling, evaluation, deployment and monitoring".
11. Design a phone assistant for a medical practice that replies within a second. Streaming speech-to-text, LLM, text-to-speech, a latency budget per component, human hand-off. Doctolib's Senior AI Engineer – Phone Assistant loop included an "AI system design" round.
12. How would you detect drift in production? Input and output distributions, proxy metrics while labels arrive, alerts, retraining triggers. Production experience appears in at least 68% of requirement lists.

LLM interview questions for AI engineer roles

AI engineer interviews add LLM systems: retrieval, evals, agents, cost and, for research roles, post-training. Of 180 ML and AI engineer requirement lists on Tokn, at least 59% mention LLMs, 34% agents and 30% evaluation. The lines read like prompts: Figma asked for "eval sets and LLMs as judges", Legora for judgment on "citations, evals and hallucination risks", Brevo for "LLM cost optimization at scale (caching, routing, model selection)". More on what an LLM / AI engineer does.

13. How do you build an eval set, and when can you trust an LLM-as-judge? Start from real failures, write a rubric, calibrate the judge against human labels, run it in CI. Hamel Husain's 2024 essay "Your AI Product Needs Evals" (summarised by Simon Willison) remains the standard read.
14. Your RAG system cites sources that don't support the answer. How do you debug it? Separate retrieval from generation failures, check recall@k and chunking, add a citation check.
15. When would you fine-tune instead of prompting or RAG? Retrieval for knowledge, fine-tuning for behaviour or format, always guarded by evals.
16. Explain the KV cache. It stores past tokens' keys and values so they aren't recomputed; its memory grows with context and batch size. Perplexity's inference ad mentioned "KV-cache management".
17. SFT, DPO and RLHF: what does each optimise? SFT imitates demonstrations; DPO learns from preference pairs without a reward model; RLHF trains a reward model, then optimises against it, which invites reward hacking. H Company's post-training ad listed all three.
18. Design a support agent that uses tools. How do you stop loops and control cost? Tool schemas, step limits, timeouts, tracing, per-task evals, a cost budget. ClickHouse asked for someone to "own the agent stack end-to-end".

Take-homes, case studies and presentations

Take-homes are less common than live rounds: 12 of the 94 technical postings that publish a process mention one, at 6 companies. Mistral AI's data scientist homework came with five days to prepare and a 60-minute presentation; Photoroom's Product ML Engineer loop had a home assignment reviewed with the team.

Present problem, baseline, evaluation, what failed and what's next. Canva's engineering blog says it assesses "what they'd improve with more time, and how their approach would scale".

19. What would you do with another week, and what breaks at ten times the load?

Past-project and behavioural questions

Most published loops include a values round, and several a deep dive on a project you led. Communication appears in at least 36% of requirement lists, ownership and ambiguity in 21%.

20. Walk us through your most impactful project. Synthesia's guide lists the follow-ups: "Why was this project a priority?… How did you define success? What metrics…? What was the end result…? What would you do differently?" Lovable's Data Scientist, Agent loop had a "Most Impressive Project" session. Bring numbers.
21. Tell me about a model that worked offline and failed in production. What you missed and what you changed.
22. How do you use AI coding tools, and how do you check their output? A Mistral AI Applied AI Software Engineer ad asked candidates to "use AI coding tools and coding agents extensively".

Can you use AI tools in an ML interview?

It depends on the company and the round, and policies change often. Ask the recruiter before each round.

We think you should practise both with and without AI, since OpenAI's guide shows one loop can mix both rules. Our comparison of AI coding tools covers the ones interviewers name.

How to prepare for an ML interview in two weeks

Week one is fundamentals and coding; week two is design, LLM systems and your stories. For the longer path, see the ML engineer roadmap.

Week 1

• Refresh fundamentals with Chip Huyen's free Introduction to Machine Learning Interviews. It dates from 2021, before LLM-era interviews: use it for fundamentals and process, not the LLM round.
• Code daily in Python, NumPy, PyTorch and SQL, on data and model code.
• Break a training loop on purpose, then debug it.

Week 2

• Do three system design mocks out loud, one of them an LLM system.
• Build one small RAG project with an eval set and a cost figure.
• Write three project stories with metrics and what you would change.
• Read the company's interview guide if it publishes one.
• For LLM rounds, read Chip Huyen's AI Engineering (December 2024).

Find machine learning jobs and set an alert

Browse machine learning jobs or LLM and GenAI engineer jobs, and set up a job alert to hear about new ones first.

FAQ

How many rounds does a machine learning interview have?

Usually four to six stages in published loops: recruiter screen, hiring manager, one to three technical rounds and a values round. OpenAI's guide, as of October 2026, describes 4–6 hours of final interviews with 4–6 people.

Is LeetCode still needed for ML interviews?

Less than for software roles. Synthesia and H Company describe their coding rounds as practical, not puzzles. Practise Python, PyTorch and SQL on realistic tasks, plus debugging.

Can I use ChatGPT, Claude or Copilot in an AI interview?

Only when the company says so. Canva expects AI tools, while Anthropic's July 2025 guidance rules them out in live interviews unless told otherwise. Ask the recruiter.

How long does an ML hiring process take?

Published targets run from 10 days (Qonto) to two or three weeks (Synthesia, Shift Technology). OpenAI's guide mentions about a week for application review and again for the final decision.

Do I need a PhD to pass an ML interview?

Not for most engineering roles. A research record (papers or a PhD) appears in at least 29% of ML and AI engineer requirement lists on Tokn, and OpenAI's guide says "We are not credential-driven." Research roles add a research discussion.