AI Engineer interview prep, what to expect

It is enough to get the screen, and whether it gets you the offer depends on what you can say about that chatbot beyond the fact that it exists. Interviewers will ask what users did with it, how you knew the answers were right, what it cost per conversation and what broke. If your honest answers are 'a few colleagues used it', 'we read some outputs' and 'we never measured cost', spend four weeks turning it into a real story before you interview: add a golden set of fifty questions with labelled answers, a nightly eval, a cost dashboard and structured output with validation. Then the same project becomes evidence of the whole loop. Your backend depth is an advantage in the build round and the design round, where most candidates from the data side are weaker on services, queues and observability. Lead with that. Do not claim fine-tuning experience you got from a tutorial; the technical screen will find it.

The work changes from training to composing. Most product teams in 2026 do not train models; they build on hosted ones, and the hard problems moved to retrieval, evaluation, orchestration, cost and guardrails. In the interview, your modelling depth still counts in the technical screen, but the design round will not ask you to design a training pipeline; it asks for a retrieval or agent system and grades the trade-offs. The evaluation discussion is where you can shine, because you already think in test sets and metrics, as long as you translate that into LLM terms: golden sets, judges, regression gates on prompt changes. The build round can surprise ML engineers who live in notebooks; it wants a small service with validation, tests and a fallback path, written the way a backend engineer writes it. Practise that. On pay, AI Engineer titles at product companies land close to MLE at the same level, and the labs pay far above both.

Three things, in this order. First, does it run and answer correctly on their documents, with a README that lets a reviewer reproduce it in five minutes. Second, how you handled the model being wrong: structured outputs with validation, citations back to the source chunks, a sensible answer when retrieval finds nothing, and tests that replay bad outputs. Third, your judgment on trade-offs, stated in the README: why this chunk size, why hybrid search or why not, what it costs per query, what you would build next with another week. Candidates lose the take-home by spending the weekend on a fancy UI and no evaluation. A small eval script with twenty labelled questions and a score is worth more than any interface. Keep the stack boring and the code tested; the review conversation afterwards is where they check you understood every line, including the ones a coding assistant wrote.

Less than you expect, as long as you can explain model behaviour from mechanism. The technical screen asks what a token is, why context windows matter, why temperature changes answers, what an embedding is for. That is a few weeks of reading, and Chip Huyen's AI Engineering covers it in the order interviewers ask. Nobody at a product company will ask you to derive attention. What they will ask is whether you have shipped something on a model, so build one real feature before you apply and measure it. Your full-stack background is an asset in the build round, where the grade is on a working service with validation and tests, and in product-sense conversations, where you can talk about what users do. The gap to close is evaluation: learn what a golden set, an LLM judge and a regression gate are, and use them on your own project so the story is yours.

The startup loop is usually shorter and more practical: a founder or lead on the screen, a build round on their real problem, sometimes a paid work trial, and a design conversation that goes deep on one system rather than surveying many. They grade speed, taste and whether you have used the current tools for real. The large company loop is longer and more standardised: a coding screen that may still include an algorithm question, a design round from a rotation of prompts, a separate behavioural round with a leveling rubric, and a hiring committee. They grade consistency and whether you fit a level. Prepare for both by having one shipped feature you can take apart from any angle, three canonical designs you can draw cold, and an evaluation story with numbers. For the large company add two weeks of algorithm practice; for the startup, use their product for a week and arrive with two specific things you would change.

No. The ML Engineer loop grades training, serving and ML system design at moderate scale, with a coding bar close to a software engineer's. The AI Engineer loop grades building on hosted foundation models: retrieval, structured outputs, agents, evaluation and cost. If the posting says train, fine-tune or serve models, read the ML Engineer guide. If it says build features on LLM APIs, RAG, agents or evals, this is the one. Many teams use the two titles loosely, so read the responsibilities in the posting rather than the title.

The process takes 3–5 weeks. Before it starts, ship one small LLM feature end to end with an evaluation script and tests, and drill three canonical designs: document Q&A over a large corpus, a tool-calling agent, and an evaluation pipeline that blocks a release. The evaluation discussion is the round most candidates have never prepared for.

Describing components instead of failures. Candidates list vector databases and re-rankers and never say what breaks, how they detect it, or what it costs. Interviewers grade the trade-offs and the evaluation story, and a design with neither sounds like a tutorial.

Most variation is at the edges. Major tech companies (FAANG, scale-ups, mid-size SaaS) follow processes within 1–2 rounds of what's described. Smaller startups often run fewer rounds (3–4) but the bar at each round is similar; less-tech-mature companies sometimes skip system design or behavioural rounds entirely. Read the posting and ask the recruiter on the screening call; they'll tell you what's coming.

This guide covers the general bar at L4 / IC3 to IC4. The free scan reads your specific job description and returns predicted questions for that exact role + company, a pay benchmark matched to the role and, with your CV, a read on your experience gaps and an ATS resume check, emailed as a PDF.