Skip to main content

How to prepare for an AI interview: practice out loud, against the clock

By Xplaind editors ·

Cover: "Answer out loud. Against the clock." beside a stopwatch drawn in teal and brass

Most candidates prepare for AI interviews by reading answers. Then, in the interview, they know the answer but cannot say it clearly in the time they have. Speaking an answer is a different skill from recognising one, and it improves only with practice out loud.

What AI interviews ask

The mix depends on the role, but most AI interviews draw from the same pool:

  • Fundamentals. How a language model produces an answer, what tokens and context windows are, why models hallucinate.
  • Retrieval and agents. How RAG works and how to debug it; how tool calling works; when an agent is the wrong choice.
  • Evaluation. How you would know a change made the system better. Strong candidates always come back to this.
  • Approach choice. Prompting, RAG or fine-tuning, and why.
  • System design. "Design a customer support assistant for an Indian bank", with trade-offs on cost, latency, safety and handoff to people.
  • Behavioural. A project you built, a time something failed, a disagreement about using AI.

Forward deployed engineers also get client scenarios: run a discovery session, handle scope creep, explain a delay to an executive. AI product managers get product sense, metrics and responsible AI questions. ML engineers get classical machine learning: overfitting, leakage, metrics for imbalanced data.

We wrote questions for each role, with short answers, deeper answers, follow-ups and common mistakes. Browse them by role: AI Engineer, Generative AI & LLM Engineer, Forward Deployed Engineer, Machine Learning Engineer, AI Product Manager, or all roles. Each role also has a Top 25 with examples you can use in the interview.

A structure for any spoken answer

Interviewers remember answers that are easy to follow. This four-part structure fits in one to two minutes:

  1. The point, in one sentence. "RAG lets the model answer from our own documents by retrieving the relevant passages first."
  2. How it works, in two or three steps. Not every detail; the steps that matter.
  3. An example or a number. From your own project if you have one.
  4. A trade-off or a risk. "It adds latency and depends on retrieval quality, so I would measure recall on a test set." This is what makes an answer sound senior.

Then stop. Let the interviewer ask the follow-up; that is a conversation, not a failure.

Why the clock matters

Real interviews have time pressure. If you take three minutes to reach your main point, the interviewer will interrupt before you get there. Practicing with a limit teaches you to lead with the point and leave out what does not help.

Practice it here

Our mock interview works like the real thing:

  • choose a role and 3, 5 or 10 questions;
  • each question shows with a timer of one to two minutes;
  • press Space to turn on the microphone and answer out loud; your words appear as text, which you can edit, or you can type instead;
  • press Enter to submit.

At the end you get a report: which key points each answer covered and missed, your speaking pace, filler words, answers that ran out of time or ended too early, and the three things to fix first, with links to the lesson for each gap.

It is honest about its limits: it checks whether you mentioned the key points, not whether your explanation was correct, so read the model answers too. Your voice and answers are not stored on our servers. Speech-to-text is done by your browser; in Chrome and Edge that sends your audio to Google or Microsoft, so type your answers if you prefer.

A two-week plan

  • Days 1 to 3: read the questions for your role and the lessons behind any you cannot answer.
  • Days 4 to 10: one five-question mock interview a day. After each, redo your weakest answer out loud using the four-part structure.
  • Days 11 to 14: ten-question sessions, plus one practice system design with a friend or mentor who interrupts with follow-ups.

Record yourself once and listen back. It is uncomfortable, and it is the fastest way to hear what an interviewer hears.

Questions people ask

What questions are asked in an AI engineer interview?
Expect questions on how language models work (tokens, context, temperature), retrieval-augmented generation, agents and tool calling, evaluation, choosing between prompting, RAG and fine-tuning, system design for an AI product, and behavioural questions about projects you built.
How long should an interview answer be?
For a concept question, about one minute. For a design or behavioural question, about two minutes before the interviewer steps in with follow-ups. Lead with the main point so you are not cut off before you make it.

Today's AI turns are used

Lessons, roadmaps, videos, practice cards and the code playground stay free with your account, with no limit.