AI Engineer Interview Questions: What 1,988 Questions Show

We sorted 1,988 AI engineer interview questions from four public sources and found two rankings, not one. In the questions people reported from real interviews, the repeats are defending a project you shipped, keeping an AI agent from failing in production, and choosing a vector index.

Study banks rank a different list first: system design, eval catalogs, serving internals, and fine-tuning. Fine-tuning shows up in 137 questions across the lists and in zero that a candidate reported. Unlike a top-50 list, this study tags every question by where it came from, so you can tell a question a room asked from one a coach wrote down.

What real AI engineer interviews ask vs. what study banks ask

Candidates who described a real interview put agent reliability, system design, and project defense on top. Study banks put system design, evals, and serving on top, and spend 119 questions on fine-tuning.

Each column counts questions from that source type. A column is not the chance a question comes up in your loop, and the columns don't add into one total.

First bar: candidate reports (44 questions). Second bar: study banks (1,154).

If your prep list came from a question bank, it leans toward the right-hand column.

Project defense: the question that opens the loop

Seven of the 45 candidate-reported questions ask you to walk through a project you owned. These are the prompts candidates wrote down:

• Explain a project you owned end to end and how you dealt with it.
• What's the last GenAI project you have worked on?
• Explain that project's architecture.
• Explain the retrieval pipeline you worked on.
• Walk through a RAG pipeline you have built.
• Walk through the RAG pipeline you built and which embedding model you used.
• Walk through the end-to-end architecture of a production system you shipped.

Three of the seven ask you to defend your choices, not just describe them. Three also come from a single post, so read this as a strong pattern, not a rate. Study banks list 21 more project-defense prompts.

Build one project story before anything else: the architecture, the retrieval pipeline, the model you picked and why, and what broke in production.

The project story to rehearse before any other topic

AI agent reliability interview questions

Agent reliability is the largest group in the candidate column: 11 questions. Interviewers want to know whether your agent holds up once it leaves the demo. The circuit-breaker question below is labeled API and limits, but it belongs in the same conversation.

Debugging and latency

• An AI agent works in demos and fails randomly in production. How do you debug it?
• A Python script calls an API in 10ms. The same call made by the LLM takes 310ms. Why?
• Why does an LLM making a tool call take longer than a normal Python API call?
• Where do circuit breakers, rate limiting, and fault tolerance sit in an agent system?

Tool calls, memory, and state

• How do you prevent hallucinations during tool calling?
• What is memory in a voice agent, if it is not the context window?
• How do you manage session state in a RAG system?
• How do LangGraph supervisor agents, memory, and checkpointers work in your design?

Frameworks and multi-agent systems

• What are MCP internals?
• Which agent framework is current, and what does it actually change?
• How do agents communicate, and when does multi-agent orchestration break?
• How does your agent workflow run?

MCP is the only tool the whole corpus names often: 23 questions, one of them from a candidate.

For every agent you've built, know one production failure, how you found it, and what you changed.

RAG, retrieval, and vector index questions

RAG is the most named concept in the corpus: 82 questions, 10 of them from candidates. The candidate questions go past "explain RAG" and into the choices you make at scale.

Vector indexes

• Which ANN approaches did you use in RAG, and why use ANN at all?
• For 50 to 100 million chunks, which index do you choose, and how do you decide?
• Draw HNSW vs. IVF. Which is better, and when?

Vector indexes run opposite to fine-tuning. The whole corpus holds only 12 index questions, and 3 of them come from candidates.

Vector indexes: 3 of 12 questions came from candidates. Fine-tuning: 0 of 137.

Chunking, embeddings, and retrieval quality

• Explain a recursive text splitter and when you use it.
• How do you handle a vector database for a document of about 3,000 pages?
• How do embeddings work, and how do you use them in a RAG system?
• How do reranking and chunking strategies change retrieval quality?
• What happens when the retrieved context is not good enough to answer the question?

Proving it works

• What evaluation metrics do you use for RAG?
• How do you know it works?

Be ready to pick an index for a stated corpus size, defend the pick, and say how you would measure whether retrieval got better.

System design and model behavior questions

System design is the second-largest candidate group, with 8 questions. Most of them are RAG systems on a real cloud, plus a few classic distributed-systems prompts.

System design

• Design an enterprise RAG assistant over internal company documents.
• Design an end-to-end AWS RAG architecture.
• Discuss RAG pipelines on AWS.
• Explain RAG architecture.
• Design an AI interview platform that conducts interviews.
• High-level design plus database design.
• Explain SAGA and two-phase commit.
• Locking and concurrency.

Model behavior and safety

• Temperature 0 should make an LLM deterministic, but two runs of the same prompt differ. Why?
• What is the lost-in-the-middle problem in a long system prompt, and what do you do about it?
• How do you prevent prompt injection, and what belongs in the system prompt?
• Name an inference optimization and when it does not help.
• Explain tokenization.

Coding and classic ML

• Explain a small gradient-descent code snippet.
• A Python coding round. The candidate did not share the prompt.

Most candidate questions still ask you to explain something: 27 of them. Only 5 ask you to design from scratch. Practice explaining each concept in two minutes, with one tradeoff you've hit yourself.

Task behind each firmly labeled candidate question. Explaining a concept dominates.

What study banks over-teach, and what company tags really mean

Four topics are large in the banks and close to absent from what candidates reported:

• Fine-tuning: 137 questions, none from a candidate.
• Attention: named in 57 questions, none from a candidate.
• Prefill and decode: named in 18 questions, all from study banks.
• Data plane (ingest and permissions): 22 questions, none from a candidate.

First bar: all sources. Second bar: reported by candidates. RAG is the only concept candidates named more than once.

Evals look big too, with 216 questions, but most ask you to list which metrics exist. Across the whole corpus, only 26 questions ask you to judge an actual output.

That doesn't mean nobody asks about fine-tuning. It means the candidates in this corpus didn't write it down. Study it after the candidate topics, not before.

Company tags are labels, not leaks

Banks often file questions under a company name. This is what those tags pile onto:

Of 1,326 company mentions in the raw data, 1,309 are tags a bank author added. OpenAI's 173 tagged questions include no candidate who said they interviewed there. The 17 mentions that came from candidates name three employers, and none of them is an AI lab.

1,309 of 1,326 company mentions are bank tags. Candidates named three employers, none of them AI labs.

Use a company tag to see what a bank author thinks that company cares about. Don't read it as what that company asked.

How to prepare for an AI engineer interview with this data

Work from the candidate column down, then fill gaps from the banks.

1. Your project story. One project, end to end: architecture, retrieval pipeline, model choices, and what broke.
2. Agent failure modes. Demo vs. production, tool-call latency, hallucinated tool calls, state and memory.
3. Retrieval at scale. Chunking a large document, picking an index for tens of millions of chunks, and measuring retrieval quality.
4. RAG system design on a real cloud. Enterprise assistant over internal docs, end to end.
5. Bank topics last. Fine-tuning, attention math, and serving internals, once the first four are solid.

Prep order, from the candidate column down to the study banks

Some production topics don't show up as named prompts at all. The 1,988 questions never mention backpressure, leases, or swarms. Idempotency appears once. Rate-limit errors (429) and human approval gates appear twice each, all in study banks. If you have shipped answers here, you bring something the lists don't cover.

Get the 4Geeks AI Engineering interview kit: Download the kit.

How we built the study and what it can't tell you

We logged every time a question was written down across four public source types, then kept track of which type each question came from.

Thirty-four of the Landed rows come from three Medium compilations by Adil Shamim. We cite them as compilations, not as interviews.

From 2,354 rows to 1,988 questions. We dropped 27 rows that weren't questions, leaving 2,327. We merged only identical wording, including 40 spelling variants. Different prompts on the same topic stayed separate, so "What metrics do you use for RAG?" and "Walk through a RAG pipeline" are two questions.

From raw rows to 1,988 unique questions. Three more share wording across a compiled report and a bank.

No question appears in all three source types.

Labels. A classifier read each question's text, without the source or company, and picked one topic and one task. We count a label only when its confidence is at least 0.6. That gave 1,519 firmly labeled questions; the other 469 stay in the dataset and out of the counts.

Level. Of the 45 first-hand reports, we coded 22 as senior and 10 as mid from the post, and 13 named no level. The banks leave level blank on every row.

Coded from each post, not a level printed by the employer. Study banks record no level at all.

What this study will not claim:

• The percent of interviews that ask any question.
• That a company asks what a bank tags with its name.
• One combined ranking of banks and candidate reports.
• A skill tree. The links between topics loop back on each other.

Forty-five candidate questions are not the whole industry. A zero in the candidate column means these posts didn't describe that prompt, not that no company asks it.

Get the data. The full question register, labels, and figures are open on GitHub under CC BY 4.0: AI engineer interview question register.