Loading…
AI engineer, machine learning engineer, ML platform and MLOps interviews at Google, Amazon, Nvidia and the AI-first companies are systems interviews about a component that happens to be a model: how you serve it, batch it, retrieve for it, keep it from silently getting worse, and keep the accelerator bill sane. This plan sequences the site's questions, guides, quizzes and concept pages into six weeks — two on the coding these loops actually ask, one on inference and serving, one on the AI design round, one on training and cost — easiest first, with a free week one so you can judge the format before paying for anything.
Built for someone with a few weeks before an AI engineer, machine learning engineer, ML platform, AI infrastructure or MLOps loop who already ships services and has used a model in production. It is not a research-interview plan: nothing here derives an architecture from first principles. If your loop is a pure software engineer loop with an ML flavour, the software engineer plan's coding weeks are the better spine.
Nine weeks, 59 steps, about 26 hours. Week one is free with no account.
Weeks 1, 2 of the six-week plan, combined.
Six weeks at roughly four to five hours a week, or about 26 hours in total. The plan compresses to as few as two weeks by dropping the optional steps first, and stretches to twelve; the core questions stay in every version.
Week one is entirely free with no account: the guide, the concept page, all three questions, the serving discussion question and the quiz. Weeks two to six use Pro questions for the coding, inference and design rounds; every blog post, concept page, quiz and tool stays free throughout.
Progress is derived from questions you have completed and quizzes you have taken, not from a stored position in the plan. If you solved the KV cache question last month, the plan already shows it done. Reading steps are not tracked, so the percentage counts questions and quizzes only.
Week one is free. From week 2 the plan uses 32 Pro questions and 2 Pro behavioral guides; the posts, concept pages, quizzes and tools stay free.
See Pro pricingWrite the retrieval end from scratch — similarity, chunking, nearest neighbours, a pipeline, rank fusion — and read the RAG guide that puts it in a system.
The coding questions these loops actually ask sit at the retrieval end: a similarity metric, a normalisation, chunking, nearest neighbours, a retrieval pipeline, and how two rankings fuse. The point is to show you know what the library is doing, not to reproduce it from memory.
Weeks 2, 3 of the six-week plan, combined.
The coding questions these loops actually ask sit at the retrieval end: a similarity metric, a normalisation, chunking, nearest neighbours, a retrieval pipeline, and how two rankings fuse. The point is to show you know what the library is doing, not to reproduce it from memory.
Only if the set above felt routine: a greedy walk through a layered graph is what an approximate nearest-neighbour index actually does, and the vector database question in week five assumes you know it.
A loss and its gradient, an activation and its derivative, a normalisation layer, the metrics you would evaluate with. Interviewers use these to check that you understand what the model computes, which is the prerequisite for noticing when it computes something wrong in production.
The numerical questions — loss, gradient, activation, normalisation, metrics — then evaluation, because noticing a model getting worse is the operational half the loop weights most.
A loss and its gradient, an activation and its derivative, a normalisation layer, the metrics you would evaluate with. Interviewers use these to check that you understand what the model computes, which is the prerequisite for noticing when it computes something wrong in production.
Weeks 3, 4 of the six-week plan, combined.
Backpropagation through one layer and scaled dot-product attention are the two things a senior AI infrastructure interviewer may still ask you to write. Once each, on paper, is enough.
The half of the loop that makes a model different from any other service: the KV cache, batching, sampling, and the platform that serves them at a cost you can defend.
Weeks 4, 5 of the six-week plan, combined.
Retrieval at scale, the vector store under it, and the feature store beside it — the design questions asked of ML platform loops, with the numbers.
The training platform under time, the cost conversation every AI infrastructure loop has, and the behavioral round that vetoes strong technical loops.
Typically a design round on an inference platform, a retrieval system, a feature store or a training platform; a discussion round on serving trade-offs, drift and cost; coding drawn from the numerical and retrieval end rather than classic algorithm puzzles; and a behavioral round. The plan gives coding two weeks, inference one, the design round one, and training with behavioral the last.
Rarely, which is why both are optional steps in week three rather than core ones. These loops assess whether you can put a model into production and keep it working — serving, scaling, monitoring, cost — far more than whether you can derive it. Understanding what the model computes is required; deriving it once on paper is enough.
A machine learning engineer loop is the same loop under a different title — follow it as is. An applied scientist loop an applied scientist loop adds a research and modelling depth this plan does not cover. If your loop is a software engineer loop with an ML flavour, follow the software engineer plan's coding weeks and add weeks four and five from here.