AI engineer interview map
The 2027 AI engineer interview map
AI engineer interviews are not one topic. They test whether you can design retrieval, prove quality with an eval, keep an agent from looping, and explain what a change costs. This page maps the loop to the sixteen topics, 51 graded challenges and five guides on LearningTo, so you can see the whole board before you start.
Last updated 2026-11-05
The loop in five stages
Foundations
What is the system actually doing when it answers?
What a strong answer does. A strong answer separates retrieval from generation, names the context budget, and explains why a confident answer can still be unsupported.
Building systems
Design retrieval, fine-tuning and agent loops for a real task.
What a strong answer does. The candidate picks a path and defends it: prompt, retrieval, or training. They name the failure modes and the eval that would catch them.
Proving quality
How do you know the change made the system better?
What a strong answer does. An eval set with human agreement, a judge calibrated against it, and a clear answer for noise, contamination and pass@k versus pass^k.
Serving and cost
What happens to latency and spend when traffic doubles?
What a strong answer does. A latency budget split across queue, prefill and decode; cache hit rate; token accounting; and a fallback ladder with a cost ceiling.
Trust and failure
Where does untrusted text enter, and what stops it?
What a strong answer does. The candidate draws the trust boundary, treats tool output and retrieved text as untrusted, narrows the tool set, and puts a human in front of consequential actions.
The topic board
16 topics, 51 graded challenges, 39 glossary terms and 5 long-form guides. Each topic below links to its walkthrough and practice page.
| Stage | Topic | Challenges | Practice |
|---|---|---|---|
| Foundations | Semantic Search & Document QA When you type a frustrated question into a help-center search box and the right article still comes back, something matc | 3 | Practice |
| Foundations | Prompting & Context Windows You have typed the same question into a chat window twice, gotten two different answers, and watched a third rewording p | 3 | Practice |
| Building systems | RAG System Design in Production Building a support chatbot over a company wiki is a weekend project; keeping it correct for thousands of users is the in | 3 | Practice |
| Building systems | Fine-Tuning: When, and Why Picture a specific morning. Your support inbox fills with replies from the AI assistant your team switched on last month | 4 | Practice |
| Building systems | AI Agents: Loops, Tools, and Stopping An ordinary AI chat answers; an AI agent acts. Give a model access to tools - a search function, a database lookup, a se | 3 | Practice |
| Proving quality | LLM Evals & Judge Calibration How do you know a change to an AI feature actually made it better, before your users find out the hard way? That questio | 4 | Practice |
| Proving quality | Serving AI Reliably The AI feature launched ten days ago and is the quietest thing you operate: the dashboard is a strip of green graphs, no | 3 | Practice |
| Serving and cost | Inference Economics & Latency Every time someone sends a message to an AI feature, the model has to run on that request, and the run has a price in do | 3 | Practice |
| Serving and cost | LLM Serving Internals Before a streamed answer reaches the screen, a scheduler admits the request into a batch of others running on rented GPU | 3 | Practice |
| Serving and cost | Design ChatGPT: Tiers & Rate Limits Design ChatGPT is the most-cited systems design question at AI companies, and it is really one question: when everyone a | 3 | Practice |
| Serving and cost | Cache-Aware Routing & Affinity The next hidden cost in a serving fleet is where each request lands. Every arrival carries a shared opening - a system l | 3 | Practice |
| Serving and cost | Quantization & the Startup Stack The second serving job is not the fleet; it is the one card, when the budget is what a startup can put on the balance sh | 3 | Practice |
| Serving and cost | Buying the Fleet The third serving job is the balance sheet: the product found users, the token demand no longer fits on a demo pod, and | 3 | Practice |
| Serving and cost | Metering & Cost Attribution The fourth serving job is the meter: the model serves traffic, the fleet is paid for, and the dollars land on an invoice | 3 | Practice |
| Serving and cost | Fleet Efficiency & Eviction When a serving pool grows past a handful of cards, someone has to judge the cards, and that job is a real interview subj | 3 | Practice |
| Trust and failure | Guardrails & Prompt Injection Picture one specific failure. A customer asks your support assistant about a refund, the assistant searches the company | 4 | Practice |
How this map was built
It is a curriculum map, not a survey. The topics are the sixteen authored AI track topics; the challenge counts come from the shipped challenge data; the guides are the five long-form guides on this site. The loop stages are a teaching frame for the judgments the track drills, not a claim about any one company's interview process.
Planning ranges, such as the ten to fourteen week roadmap, are estimates and labelled as such. Company-specific loops change; confirm the format with your recruiter.
Start here
Questions and answers
What topics does an AI engineer interview cover in 2027?
The recurring set is retrieval and RAG, prompting and context limits, evaluation and judge calibration, agent loops and tool use, serving latency and cost, guardrails and prompt injection, and enough model training knowledge to know when fine-tuning is the wrong tool.
Do AI engineer interviews still test coding?
Most loops still include a coding round, usually Python and often a data or API task rather than a pure algorithms puzzle. The SWE foundations track covers the coding round; the AI track covers the system judgment around it.
How long should I plan to prepare?
The roadmap guide uses a ten to fourteen week planning range for someone with software experience and no production AI work. That is a planning estimate, not a measured outcome; the right length depends on the roles you target and how much of the loop you have seen before.
Is this map based on real interviews?
It is built from the authored AI curriculum, question banks and spec research on LearningTo, not from a survey of companies. It describes the topics and judgments the track teaches and the questions it prepares you to answer. Treat company-specific loops as something to verify with your recruiter.
What should I study first?
Start with semantic search and RAG so you can reason about evidence, then evals so you can tell whether a change helped. After that, agents and serving economics are the two areas that separate candidates in senior loops.