AI Engineering interview preparation.
Turn your understanding into examples, reasoning, and clear trade-offs.
Explain before you reveal.
Choose a level, explore explained answers, then try a focused session.
84 questions
What is an LLM good for?
An LLM processes language using learned patterns and supplied context. I use it to summarize, extract, classify, explain, and propose tool actions. Current claim status and business permissions come from services, not the model's memory.
Does temperature zero guarantee the same correct result?
No. It generally reduces sampling variation where supported, but it does not verify evidence or eliminate all runtime variation. I use schemas, authoritative data, bounded outputs, and evaluations to establish application reliability.
How do you choose a model?
I compare candidate models on representative tasks, including failure and adversarial cases. I measure correctness, supported tool/schema capabilities, latency, cost per successful task, and deployment/data restrictions, then version the selected…
The model invents a claim status. What do you change?
I remove any path that asks it to know account facts. An authorized tool returns canonical status; a template or model explains that result. I test missing claims, denied access, backend failure, and unsupported extra statements.
What makes a good prompt?
A clear task, trusted constraints, explicitly separated input, expected output, and a defined response when information is missing. For policy answers, I require supporting evidence and forbid inventing unsupported policy conditions.
Zero-shot or few-shot?
I begin with clear zero-shot instructions. If errors concern category boundaries or formatting, I add a few representative reviewed examples and evaluate them on separate cases. Examples should cover ambiguity, not just easy successes.
How do you maintain prompts?
I treat prompts as versioned configuration with an owner, evaluation dataset, release gate, and trace tag. I investigate patterns in failures instead of accumulating conflicting instructions for individual anecdotes. Provider/model upgrades rerun the same…
Adding many examples improved a demo but slowed production. What next?
I measure which examples actually help and remove redundancy. I test a smaller prompt, provider-supported schemas, deterministic preprocessing, or a better model. I optimize latency and cost while preserving held-out quality.
What is RAG?
RAG retrieves relevant external evidence and gives it to a model before answering. It lets the insurance assistant use current private policy documents and produce source-backed explanations instead of relying only on learned knowledge.
RAG or long context?
If the task already identifies a small permitted document set, long context may be simpler. For a large changing corpus, retrieval selects relevant evidence more efficiently. Both need access control, applicability checks, and answer evaluation.
How do you debug wrong RAG answers?
I inspect parsing, document version and permissions, chunk boundaries, candidate recall, ranking, supplied context, and answer support in that order. I score retrieval and generation separately, then improve the failing stage rather than automatically using a…
An answer cites a real document but is wrong. Why?
The cited document may be irrelevant, expired, inapplicable, or missing an exception; the answer may also misinterpret it. I validate source eligibility and claim support, not just whether the link exists.
Explain an embedding.
It is a numerical vector that represents useful features of content, often meaning. Similar wording or meaning tends to produce nearby vectors, enabling retrieval of “vehicle collision” for a query about “car accident.”
What does cosine similarity tell you?
It compares vector direction and helps rank related content. A high score is not a probability that the passage is true or sufficient. I use a compatible metric and calibrate thresholds on domain-specific queries.
How do you upgrade embedding models?
I version models and index data, re-embed the intended corpus into a separate compatible space, evaluate filtered retrieval, and switch query/index routing together. I preserve source data and support rollback instead of mixing arbitrary old and new vectors.
Difficulty is editorial guidance. These questions are not a ranking of interview frequency.
All AI Engineering interview questions by concept
Use the filters above for a practice selection, or follow the full question directory below. Each answer includes its source and explanation; difficulty labels are editorial guidance.