Machine Learning Engineer interview loops

A machine learning engineer (ML engineer) is a software engineer who trains models on a company's own data and keeps them working in production. Where an AI engineer builds on models as someone else trained them, an ML engineer trains them, from scratch or by fine-tuning someone else's.

As of October 2026 the title covers two kinds of work, and a third sits beside it. The first is product models: a Pinterest posting has its ML engineer working with the home feed, ads, shopping and search teams to improve their models (Pinterest), and an Airbnb payments posting has its ML engineer training the models behind its fraud defences and personalised checkout, and keeping them running with automated retraining (Airbnb). The second is fine-tuning large language models: another Airbnb posting asks for an ML engineer to fine-tune them for products such as an AI assistant and agents (Airbnb). Beside both is the platform ML engineers train and serve on, which one Pinterest posting puts under a software engineer title instead (Pinterest). The loop below is written for the first: an engineer who owns a model making decisions in a product, from the data it learns from to the retrain that replaces it.

Nobody should be hired into the role for knowing which algorithm to reach for. The work is a model that is a little wrong about everyone and very wrong about some people, wired into a product that acts on its scores before anyone has checked them. The job is to know how wrong, to keep the offline world and the online world telling the same story, and to make the model's mistakes reversible. That is ordinary engineering discipline applied to a component whose correctness cannot be read from its code, and it is what the rounds below are written to read.

Two things make the role hard to interview for. The first is that the textbook material is abundant and fluent, so a round built on definitions measures a course, not a career. The second is that the failures worth hiring against are all disagreements between two things that should match: a feature computed one way in training and another way at request time, an offline metric that rises while the product gets worse, a retrained model whose scores move without anyone downstream being told. None of those is visible in a whiteboard answer, so each round here hands the candidate an artifact — a repository, a data pack, an experiment readout, an incident ticket — and watches what they notice in it.

Employers draw the line on AI in the interview in different places. Pinterest, whose ML postings link to its policy, invites AI in designated parts of the process, such as take-home assignments or a section of a live technical screen, where it looks at how candidates prompt, check the output and own the result; live interview questions are without AI unless the interviewer says otherwise (Pinterest, as of October 2026). Canva wrote in June 2025 that it expects machine learning candidates, as well as backend and frontend ones, to use AI coding tools in its technical interviews (Canva). None of the plans below says either way. Three of the four open on something the candidate has to notice in an artifact before anyone points to it, which an assistant reading the same artifact could point to for them; the rest of each round is code, a design or a decision defended live. The modelling and experiment reviews send their packs a day ahead, so decide before the first pack goes out, and tell the candidate which rounds, and which packs, an assistant may be used on.

A loop is a set of round plans in order. Each round scores one to three competencies, every must-have is scored in at least two rounds, and the debrief is held competency by competency rather than round by round, so no single interview decides the hire. How these plans work covers the rest of the method.

Further reading, as of October 2026. Chip Huyen's Designing Machine Learning Systems (O'Reilly, May 2022) is a book-length treatment of the job, through deployment, distribution shift and continual learning. Martin Zinkevich's Rules of Machine Learning (Google) gives nine of its 43 rules to training-serving skew, the subject of the coding round. Kohavi, Tang and Xu's Trustworthy Online Controlled Experiments (Cambridge University Press, 2020) covers the online experiments the experiment review is about. Sculley and colleagues' Hidden Technical Debt in Machine Learning Systems (NIPS 2015) names undeclared consumers and hidden feedback loops among its risks, and the retrain round has a version of each: a discount rule that relied on the scores without a contract, and discounts that change who churns.