← All posts
general

Machine Learning Engineer Interview: Rounds and Prep

A machine learning engineer interview is two interviews stacked on each other. You sit the software engineer's coding rounds at close to the same bar, and then a second set of rounds on modelling that a software engineer never sees. Candidates from a data science background tend to underprepare the first half, candidates from backend engineering the second, and the loop is designed so that either gap is enough to fail.

The other difficulty is that "ML engineer" covers at least three jobs: training and improving models, building the platform that trains and serves them, and applying models inside a product. The rounds are similar across the three and the emphasis is not. This guide goes through the loop, each round, how LLM topics are changing it, twenty-five example questions, and a six-week plan.

The loop

After a recruiter call and one technical screen, most loops run three to five interviews. Two published accounts, both prep guides built from candidate reports and read in November 2026, show how much the arrangement varies.

For Meta, one guide describes a 45-minute technical screen with two data structure and algorithm problems, then a final loop for generalist ML engineers of a standard coding round (45 minutes), an AI-enabled coding round (60 minutes, a multi-file project with a built-in assistant), an ML system design round of about 45 minutes, and a behavioural round. Specialised tracks reportedly add an ML coding round with from-scratch implementations.

For Google, the same site's guide describes an optional online assessment, one or two phone screens, and an onsite of three or four technical interviews, with a round it calls ML Domain that covers concepts, your past work and how you explain ML solutions to different audiences. It names no separate ML system design round.

So expect some mix of five things: coding, fundamentals, ML system design, a project deep dive and behavioural. Ask the recruiter which of them you will get and whether fundamentals are a round of their own or folded into the others. Ask also about AI tools in the coding rounds, because the answer now differs by company; the guide to AI in coding interviews has the detail.

Coding

The coding round is the one a software engineer would recognise: arrays, hash maps, trees, graphs, heaps, often two problems in 45 minutes. Being an ML specialist buys no discount on it. If you have spent three years in notebooks, this is where to start, because it is the slowest skill to rebuild.

Some loops add ML-flavoured coding. Typical shapes, invented here for illustration: implement k-means or k-nearest neighbours from scratch, write the forward pass and gradient update for logistic regression with only an array library, compute precision and recall at k from two lists, sample from a categorical distribution, or write a batching iterator for sequences of uneven length. These reward knowing the algorithm well enough to write it without a framework, and care with shapes and edge cases.

Write Python unless told otherwise, say your approach before you type, and test with a small example by hand.

Fundamentals

Whether as its own round or as follow-up questions inside another, you will be asked to explain how models learn and fail. Derivations on a whiteboard are rare outside research roles. What is tested is whether you can reason from first principles about a symptom.

The areas that recur:

  • Generalisation. Bias and variance, overfitting, regularisation, what a gap between training and validation loss tells you, and what a missing gap tells you.
  • Evaluation. Precision, recall, the area under the ROC and precision-recall curves, calibration, and which of them you would trust with one positive in a thousand examples.
  • Data. Leakage, label noise, class imbalance, how to split data that has a time order or repeated users.
  • Optimisation. Gradient descent and its variants, learning rate, batch size, why training diverges.
  • Model families. Linear models, tree ensembles, neural networks, and an honest view of when gradient-boosted trees beat a deep model on tabular data.
  • Representation. Embeddings, feature crosses, handling categorical values with millions of levels.
  • Experimentation. A/B tests, significance, novelty effects, and why an offline gain often shrinks online.

A useful drill is to take each area and prepare one failure you have seen yourself. "Our validation score was excellent because a feature encoded the label" is remembered; a definition is not.

ML system design

This is the least structured round in the loop, and the one that sets level. The prompt is a product problem: recommend videos, rank a feed, detect fraudulent payments, predict delivery time. The Meta guide cited above says recommendation and ranking prompts are the most common there.

The mistake is to start at the model. A sequence that holds up:

  1. Frame the problem. What the business wants, and what you will predict to serve it. "Increase watch time" is a goal; "probability that this user watches more than thirty seconds of this video" is a prediction target. Say what the choice of target will distort.
  2. Data and labels. Where labels come from, how late they arrive, what bias is baked in. A fraud label may take weeks to confirm, and a recommender only observes feedback on what it showed.
  3. Features. User, item and context features; which are computed in batch and which must be fresh; how you guarantee the same value at training and at serving.
  4. Model. A simple baseline first, then the model you would ship and why. For recommendation at scale, usually two stages: cheap candidate retrieval from millions of items, then a heavier ranker over hundreds.
  5. Offline evaluation. The metric, the split, and what would make you distrust a good number.
  6. Serving. Latency budget, batch against real-time scoring, caching, what happens when a feature is missing.
  7. Online evaluation and monitoring. The experiment, the guardrail metrics, drift detection, the retraining schedule and the feedback loop the model creates.

The three classic prompts stress different steps. Recommendation is about retrieval at scale and the cold start for new users and items. Feed ranking is about choosing and weighting several objectives. Fraud is about extreme imbalance, delayed and adversarial labels, and the cost of a false positive to a real customer. Practise one of each. The older post on ML and AI system architectural patterns is a useful map of the components.

Senior candidates are separated in steps 1, 2 and 7. Anyone can name a two-tower model. Fewer can explain how the training data was shaped by the previous model, and what they would do about it.

The project deep dive

One interviewer, often the hiring manager, will ask you to walk through an ML project of your own for most of an hour. It is the only round where you choose the subject, and a surprising number of candidates arrive without having chosen.

Pick a project where you made the modelling decisions and where something went wrong. Prepare it as a narrative with numbers: the problem and why it mattered, the baseline, what you tried in order, what each attempt did to the metric, what shipped, and what happened online. Then prepare for the questions that follow: why not a simpler model, how you knew the gain was real, what the model was bad at, what you would do with another quarter.

Be exact about your own part. "We trained" invites "what did you do", and interviewers at this level can tell within a few questions whether you ran the experiments or attended the meetings.

LLM topics in the loop

Classic ML loops are absorbing questions about large language models, unevenly. The Meta guide reports that some ML system design rounds now centre on generative AI systems such as retrieval-augmented generation, advises preparing at least one LLM design, and says the behavioural round asks how you use AI in your work. The Google guide mentions generative AI only as a preferred skill for certain roles. Your own loop will sit somewhere between, depending on the team.

A sensible minimum for any ML engineer in late 2026:

  • Explain transformers and attention at the level of a whiteboard sketch, and what tokens, context windows and embeddings are.
  • Know when to prompt, when to retrieve and when to fine-tune, and what each costs.
  • Be able to design a retrieval-augmented system and say how you would evaluate it; the LLM system design interview guide has a worked answer.
  • Have a view on evaluating generated output, which is harder than classification and where classic ML judgment transfers best.

If the job description is mainly about building on hosted models, you are looking at a different loop, covered in the AI engineer interview guide.

Twenty-five example questions

These are examples written for practice, not a leaked list from any company.

Coding

  1. Given a stream of numbers, return the median after each insertion.
  2. Implement k-nearest neighbour search for a query vector, then discuss how to make it fast for ten million vectors.
  3. Write logistic regression training with gradient descent using only an array library.
  4. Compute recall at k and mean reciprocal rank for a list of ranked results.
  5. Merge overlapping time intervals from an unsorted log.

Fundamentals

  1. Training loss keeps falling and validation loss rises. What is happening, and what do you try first?
  2. Your classifier has 99 per cent accuracy on a fraud dataset. Why is that not good news?
  3. How does L1 regularisation differ from L2 in its effect on the weights?
  4. How would you split data for a model that predicts next week's demand?
  5. An offline metric improved and the A/B test is flat. Give three reasons.

ML system design

  1. Design the recommendation system for a short-video feed.
  2. Design a model that ranks search results for an online shop.
  3. Design real-time fraud detection for card payments.
  4. Design a system that predicts delivery time for a food order.
  5. A new item has no interaction history. How does your recommender treat it?

Project deep dive

  1. Walk me through a model you took to production. What was the baseline?
  2. Which experiment in that project surprised you?
  3. How did you know the improvement was real and not noise?
  4. What was the model worst at, and who was affected?
  5. What broke after launch?

LLMs and judgment

  1. When would you fine-tune a language model instead of using retrieval?
  2. How would you evaluate a summarisation feature?
  3. A stakeholder wants to replace a working classifier with an LLM. How do you decide?
  4. Your model's performance decays over three months. How do you find the cause?
  5. Tell me about a time you chose a simpler model than you could have built.

A six-week plan

For an engineer with ML experience, at eight to ten hours a week. If your coding is current, compress weeks 1 and 2.

  • Weeks 1 and 2: coding. Forty-five minutes a day on data structure problems, timed, speaking aloud. Add the five ML-flavoured implementations listed in the coding section, written without a framework.
  • Week 3: fundamentals. One area a day from the fundamentals list. For each, write a half-page explanation and one failure from your own work. Answer questions 6 to 10 aloud.
  • Week 4: ML system design. Three full designs in the seven-step order: one recommender, one ranker, one fraud system. Forty-five minutes each, on paper or a whiteboard, recorded. Review where you ran out of time.
  • Week 5: the deep dive and LLM topics. Write your project narrative with numbers and rehearse it to twenty minutes. Prepare one LLM design. Draft four behavioural stories, at least one about a model that failed.
  • Week 6: mocks. Two mock ML design rounds and one mock coding round with another person. Rework whichever went worst.

The design round is the one you cannot grade yourself, because the quality of the answer shows in how it holds up when someone questions the labels, the metric and the feedback loop. A mock with a coach who works on ML and LLM systems in week 6, or earlier if design is your weak round, tells you what a recording cannot.

Where coaching fits

ML design is a spoken skill learned by being interrupted, so one or two mock design rounds with a practitioner are worth more than another week of reading. On techinterview.coach, as of November 2026, ten of the 40 coaches list artificial intelligence as a specialty, four list generative AI and LLMs and two list machine learning; the generative AI and LLM coaches page is the place to start, and a profile will tell you whether a coach's background is in classic ML, LLM systems or both. For the coding rounds, the practice interview has a coding talk-through round with an AI interviewer and a written recap. It is free; sign in to start.

Related guides