AI interview prep for engineers
AI interview prep in 2026 means working out which rounds your specific loop contains, then preparing for those and skipping the rest. Most AI and LLM engineering loops draw from the same eight round types: recruiter screen, practical coding, AI or LLM system design, project deep dive, evaluation and reliability, agent or take-home exercise, classical ML, and behavioral. Which ones you get depends on the job, the level, and the team — not on the job title.
This hub covers engineering roles that build with models: applied AI, LLM, GenAI, agent, ML platform, and software engineers moving into AI. It does not cover research scientist loops, which are a different interview entirely.
Why preparing for an AI interview is harder than it should be
The titles are new and the loops are not standardised. Two companies can both advertise a Senior AI Engineer role and run interviews with almost nothing in common: one is a distributed systems loop with a model in the diagram, the other is a product judgment loop with a take-home agent exercise attached. The job title tells you very little about the interview behind it.
The round names are worse. "AI Design", "LLM Design", "Applied AI Deep Dive" and "Technical Design" are used interchangeably by different companies for genuinely different exercises. A candidate who prepares transformer internals for a round that turns out to be product scoping has wasted the week, and a candidate who prepares product scoping for an agent architecture round has wasted it in the other direction.
The result is that competent engineers over-prepare in breadth and under-prepare in depth. They read everything, practise nothing to the point of fluency, and walk into a round they could have owned with four hours of focused work. The fix is not more material. The fix is narrowing the target before you start.
The eight round types, and what each one actually tests
Nearly every AI engineering loop is assembled from these. Read the column on the right as the signal the interviewer is scoring, not the topic on the agenda.
| Round | Usual format | What it tests |
|---|---|---|
| Recruiter screen | 20–30 min call | Whether your experience matches the level, and what you will be told about the rest of the loop. This is your best source of confirmed information. |
| Practical coding | 45–60 min, shared editor or repo | Reading unfamiliar code, changing behaviour safely, testing, and narrating a debug loop. Increasingly with an assistant allowed. |
| AI / LLM system design | 45–60 min whiteboard | Designing a system where a model is one unreliable dependency: retrieval, serving, evaluation, latency, cost, fallbacks. |
| Project deep dive | 45–60 min conversation | Whether you personally did the thing on your CV. Interviewers push until they find the edge of your knowledge. |
| Evaluation & reliability | 45 min, sometimes folded into design | How you know a probabilistic system works before users tell you. Eval sets, metrics, regression gates, monitoring. |
| Agent or take-home exercise | 2–4 hours async, plus review | Scoping, finishing, and judgment about what to leave out. The review conversation usually carries more weight than the code. |
| Classical ML | 45 min | Framing, features, metrics, and overfitting. Common in ML engineer loops, much rarer in applied AI and LLM loops. |
| Behavioral | 45 min | Ownership, conflict, and decision-making at the claimed level. For senior roles this round is frequently decisive. |
No company runs all eight. A typical loop is four to five of them, and the recruiter screen tells you which.
What to ask your recruiter
Before you prepare anything, send one short email: how many rounds, what each one is called, how long each runs, whether any of them is a take-home, and whether an AI assistant is allowed in the coding round. Recruiters answer this routinely. Anything they confirm outranks every pattern on this page, because it is first-party evidence about your loop rather than a general tendency across companies.
Start with the guide for your role
Role titles overlap, so pick the one whose description matches the job posting rather than the one on your own CV.
- AI engineer
- Product-facing engineering with models as a component. Expect a conventional coding round plus AI system design. Full guide: the AI engineer interview.
- LLM engineer / GenAI engineer
- Closer to the model: serving, retrieval, prompting and orchestration, evaluation, and the cost and latency envelope. Full guide: the LLM engineer interview.
- Applied AI engineer
- Judgment about whether AI belongs in the product at all, plus shipping and measuring it. Full guide: the applied AI engineer interview.
- Agent / agentic AI engineer
- Planning, tool use, state, recovery, and observability for systems that act. Full guide: the agentic AI interview.
Then read the guide for the rounds you are getting
Each of these decodes one round and tells you what to prepare for it specifically.
- AI system design
- The broad framework for designing any AI-containing system. Start here if your round name is generic.
- LLM system design
- The language-model-specific cut: retrieval architecture, context construction, token budgets, failure handling.
- AI design round
- An overloaded name that can mean product scoping, applied ML design, or agent architecture. This guide helps you work out which.
- AI-assisted coding
- A working session where an assistant is allowed, required, or banned. The rules change how you prepare.
- LLM interview questions
- Nine themes interviewers probe, and why memorising answers fails under follow-up.
- RAG interview questions
- Retrieval depth: chunking, hybrid search, reranking, freshness, access control, and failure analysis.
- LLM evaluation
- Golden sets, task metrics, judge models, online monitoring, and agent evaluation.
If you have one week and no confirmed round names
This is the default allocation when the recruiter has told you nothing. Any confirmation you receive should override it immediately.
- One AI system you shipped, drawable end to end in five minutes.
- One evaluation story with a baseline, a metric, a discovered failure, and a measured result.
- Your real numbers for one system: p95 latency, cost per request, what dominates each.
- Two production failures you personally debugged, and how you detected them.
- Retrieval design trade-offs: chunking, hybrid search, reranking.
- Standard distributed systems fundamentals — they do not go away.
- Two or three behavioral stories at the level you are interviewing for.
- Transformer architecture derivations, unless the role is research or inference-kernel work.
- Model release trivia and benchmark leaderboards.
- Grinding algorithm puzzles, unless the recruiter confirmed a puzzle-format round.
A seven-day plan that survives contact with a real loop
Front-load the work that pays off in every round, and leave the round-specific work until you know the round names.
Days 1–2 — Fix your evidence
- Write the end-to-end description of one AI system you shipped. Draw it. Time yourself at five minutes.
- Collect your real numbers. If you do not have them, say so in the interview rather than inventing them.
- Email the recruiter the round-structure question. Everything below depends on the answer.
Days 3–4 — The design round
- Work two design prompts end to end out loud, from requirements to evaluation to failure handling.
- Practise stating trade-offs explicitly: quality against latency, cost, and risk.
- Prepare the answer to "how would you know this is working?" before you are asked it.
Days 5–6 — Coding and deep dive
- One timed session in an unfamiliar open-source repo: fix a real issue with a test, in 45 minutes.
- Rehearse the project deep dive against a friend who is instructed to keep asking why.
- Write down the three questions you would least like to be asked, and answer them.
Day 7 — Narrow, do not broaden
- Re-read the job description and mark every requirement you cannot evidence. That is your exposure.
- Prepare your questions for them. Vague answers about evaluation practice tell you a lot about the team.
- Stop adding topics. Depth on four things beats recognition of forty.
Common failure modes in AI engineering loops
These are patterns, not certainties. They recur often enough to be worth naming.
- Designing the happy path beautifully and having nothing to say when retrieval returns junk or the model is confidently wrong.
- Answering "how would you evaluate this?" with a benchmark name instead of an eval set, a metric, and a gate.
- Claiming a system on the CV that the candidate reviewed rather than built. Deep dives find this quickly.
- Treating the model as the whole design and skipping ingestion, serving, caching, and observability.
- Being unable to say which part of a design you would cut first under a latency or cost constraint.
- Reaching for a model where rules, a lookup, or plain retrieval would be correct — and not acknowledging it.
What this site will and will not tell you
No legitimate service knows the questions a specific company will ask. Anything advertising leaked or exact questions is either selling stale forum scrapes or making it up. These guides describe patterns that hold across many loops, and they say so.
What can be done honestly is prediction with visible confidence. Something your recruiter told you is confirmed. Something inferred from the job description and recent public reports is inference, and should be labelled as such. Something with no usable signal is unknown, and hiding that would make a map look better and be worth less.
That distinction is the entire product. The guides here are the general layer; the map is the layer that applies it to your job, your level, and your recruiter's exact wording.
Common questions
- How long should I prepare for an AI engineering interview?
- Most candidates with a scheduled loop have between three and fourteen days, and that is enough if the target is narrow. The constraint is rarely knowledge; it is that unfocused preparation spreads thin across rounds that may not exist in your loop.
- Do I still need to practise algorithm questions?
- It depends on the company, and your recruiter can tell you. Many AI engineering loops have replaced the puzzle round with a practical coding session in a real repository. Some have not. Ask rather than assume in either direction.
- Is an AI engineer interview different from a machine learning engineer interview?
- Usually yes. ML engineer loops more often retain a classical ML round covering framing, features, and metrics. AI and LLM engineering loops more often replace it with retrieval, evaluation, and production reliability content.
- What if the job description is vague?
- Then the recruiter's wording becomes the strongest available signal, followed by the team's public engineering writing. A vague posting is itself information: it often means the team is still defining the role, and scoping judgment will be tested.
- Can you tell me what a specific company asks?
- No, and neither can anyone else honestly. What we can do is predict the likely shape of a loop from the job, the level, and what your recruiter said, then show why each round is ranked where it is and mark the parts that are unknown.
Related interview guides
- AI engineer interview: rounds and prep
The generalist AI engineering loop: coding, AI system design, project deep dive, behavioral.
- LLM engineer interview: rounds and prep
Serving, retrieval, agents, evaluation, and the latency and cost judgment this title is screened on.
- AI system design interview: framework and prep
A reusable framework for designing systems where a model is one dependency among many.
- LLM interview questions, organised by what they test
Nine themes instead of a question bank, and why memorised answers fall apart under follow-up.
How every ranking on this site is reasoned about, including what counts as evidence: read the methodology.
Map your actual interview.
This page is the general pattern. Paste the job you are actually interviewing for and the wording your recruiter used, and the map ranks the rounds you are likely to face, explains why each is ranked where it is, and tells you what to skip.
See a complete Full Interview Map example