yoklainterview sim
ProgressHistory

AI Engineer Interview Questions

2025 verified AI Engineer interview questions — solve with answers, learn from explanations, test yourself in a real simulation.

Try the real simulation →

Sample questions

Agents Tool UseDifficulty 1
When an LLM performs 'tool calling' during a conversation, what actually happens?
  • aThe model generates a structured request (a tool name and arguments) describing which tool it wants to use; the calling application executes the tool and returns the result.✓
  • bThe model directly executes the tool's code inside its own runtime and returns the result to the conversation.
  • cThe model pauses the conversation and executes the tool on the provider's servers automatically, without any code from the application.
  • dThe model calls the tool's network endpoint by itself, bypassing the application entirely.
Explanation:The model never executes code or makes network calls itself — it only produces a structured intent (tool name + arguments). The host application is responsible for parsing that intent, running the actual tool, and feeding the result back. Options b, c, and d all assume the model has execution capability it does not have.
Agents Tool UseDifficulty 1
Why does the description text in a tool's schema matter for tool calling?
  • aIt is only used for human documentation and has no real effect on the model's behavior, since the model decides which tool to call purely from the function's internal implementation code, not from schema text.
  • bThe model uses the tool's name, description, and parameters to decide whether and how to call it, so a vague description raises the chance of wrong calls.✓
  • cIt determines the order in which tools execute on the server, regardless of what the model requests.
  • dIt is required only for tools that return JSON; tools returning plain text do not need a description.
Explanation:The schema (name, description, parameter definitions) is the model's only source of information about what a tool does and when to use it. A poor description makes the model more likely to pick the wrong tool or fill in arguments incorrectly. It is not purely documentation — it directly shapes model behavior.
Agents Tool UseDifficulty 1
In the classic 'agent loop' pattern, what is the typical order of steps within one iteration?
  • aAct first, then think, and repeat without ever looking at the result.
  • bObserve the environment once at the very start, then keep repeating that same initial action regardless of what any later observation actually shows.
  • cThink about the next step, take an action such as a tool call, observe the result, and use that observation to decide what to do next.✓
  • dCall every available tool simultaneously, then think about which of the results to keep.
Explanation:The standard agent loop is think → act → observe → repeat: the model reasons about the next step, takes an action, receives the observation (tool result), and folds that observation into the next reasoning step. Options a, b, and d skip or misorder the feedback from observation to reasoning.
Agents Tool UseDifficulty 1
What does the ReAct pattern in LLM agents refer to?
  • aInterleaving explicit reasoning steps (thoughts) with actions such as tool calls, so each action is guided by reasoning and each result updates it.✓
  • bA UI rendering framework used by some frontend applications to display an agent's chat interface, unrelated to how the model itself reasons or calls tools.
  • cA technique where the model only ever reacts passively to user messages and is never allowed to initiate a tool call on its own during the conversation.
  • dA caching strategy where the application reuses a previous tool's output instead of calling the tool again, regardless of whether the underlying data may have changed.
Explanation:ReAct (Reason+Act) is a prompting/agent pattern where the model alternates between reasoning traces and tool actions, using each observation to refine subsequent reasoning. It is unrelated to the JavaScript UI library sharing a similar name, and it does not describe caching or a tool-free chat mode.
Agents Tool UseDifficulty 1
During a conversation, the model produced the following structured output:
{
  "tool_call": {
    "name": "get_weather",
    "arguments": {"city": "Ankara"}
  }
}

What should the application do next?
  • aIgnore the structured output and continue generating a final answer directly, since the model already produced this as if it already knew the current weather in Ankara.
  • bSend this raw JSON back to the user as the final response.
  • cStore it in a database as a completed action log without running anything.
  • dParse the arguments, run get_weather with city="Ankara", and return the result to the model as the tool's output so it can continue.✓
Explanation:This JSON is only a request to call a function — the model has not obtained any weather data yet. The application must run the actual function with the given arguments and feed the result back as an observation. Options a, b, and c either skip execution entirely or expose an internal request to the end user.
Agents Tool UseDifficulty 1
A support feature needs to answer 'What is our refund policy?' by looking up a single static FAQ document that rarely changes. Which approach fits best?
  • aBuild a multi-step agent with a tool-calling loop, step limits, and retry logic so it can search for the answer.
  • bA single prompt containing the relevant FAQ content is enough; there is no need for an iterative agent loop for this fixed, single-step lookup.✓
  • cAn agent with dozens of tools is required, because any customer question could theoretically need several tool calls.
  • dThe model must be given open-ended, always-on web search access so it can independently re-verify the refund policy from scratch on every single request, even for this fixed document.
Explanation:When the needed information is already known and fixed, a single prompt (with the content included or one direct lookup) solves the task without the overhead of an iterative loop, step limits, or retries. Building a full agent for a fixed, single-step answer adds complexity without benefit.

Topics

Agents Tool Use 75Aimlops Deployment Strategies Ml 75Aimlops Experiment Tracking Reproducibility 75Aimlops Feature Store Data Versioning 75Aimlops Ml Cicd Pipelines 75Aimlops Model Monitoring Drift Detection 75Aimlops Model Versioning Registry 75Aipy Async Concurrency Ml Serving 75Aipy Data Pipeline Dataloader 75Aipy Gpu Memory Management 75Aipy Numpy Vectorization Broadcasting 75Aipy Python Ml Packaging Environments 75Aipy Tensor Ops Autograd 75Embeddings Vector Search 75Evaluation Testing 75Fine Tuning Adaptation 75Inference Serving 75Llm Fundamentals 75Prompt Engineering 75Rag Chunking Indexing 75Rag Embeddings Similarity 75Rag Evaluation 75Rag Generation Context 75Rag Reranking Fusion 75Rag Retrieval 75Rag Retrieval Search 75Safety Guardrails 75

By specialization

LLM/RAG 450MLOps 450Python 450

By seniority

AI Engineer JuniorAI Engineer MidAI Engineer Senior

Other fields

BackendFrontendFullstackDevOps / CloudDatabaseMobileSecurityQA / Test AutomationData EngineerGame DevelopmentML Engineer

Test yourself against the 2025-question AI Engineer bank.

Start interview