Technical practice
How would you design a multi-step agent workflow with conditional loops?
What to coverThe posting calls for agentic graph architecture, state management and dynamic loops. Suggested answer focus: explain how you would represent steps and transitions, preserve state, and prevent a loop from continuing indefinitely.
How would you make tool calls and structured outputs reliable?
What to coverTool and function calling, structured outputs and error handling are stated requirements. Suggested answer focus: discuss validating tool inputs and outputs, handling failures, and deciding when the agent should retry or stop.
How would you manage context and memory across an agent workflow?
What to coverThe role explicitly includes context management and memory systems. Suggested answer focus: distinguish information needed for the current step from information retained across steps, and explain the trade-offs in relevance, latency and cost.
How would you improve a prompt that works in a prototype but fails inconsistently in production?
What to coverPrompt engineering and the transition from prototype to production are part of the posting. Suggested answer focus: describe how you would isolate failure cases, revise instructions or output constraints, and check improvements against representative examples.
How would you design a RAG pipeline using embeddings and a vector database?
What to coverThe posting names RAG, embeddings, vector search and vector databases such as pgvector, Pinecone, Weaviate and Qdrant. Suggested answer focus: walk through indexing and retrieval choices, then explain how you would check whether the retrieved material supports the generated answer.
How would you diagnose poor retrieval and decide whether to add reranking?
What to coverRetrieval optimisation and intelligent reranking are stated responsibilities. Suggested answer focus: separate failures to find relevant material from failures to rank it well, and weigh relevance gains against added latency and cost.
How would you evaluate an agent before and after a change?
What to coverThe role calls for evaluation of accuracy, reliability, latency and cost efficiency; benchmarking and automated testing are preferred. Suggested answer focus: define representative tasks and failure cases, measure each dimension, and explain how you would judge trade-offs rather than relying on a single score.
What would you observe and guard against in a production LLM application?
What to coverProduction observability, guardrails, logging and error handling are explicit responsibilities. Suggested answer focus: identify useful workflow and tool-call signals, explain how failures would be surfaced, and describe how guardrails could be tested without assuming they eliminate risk.
How would you structure a Python REST service that runs agent workflows asynchronously?
What to coverThe posting requires Python, OOP, asynchronous programming, REST APIs and backend service architecture. Suggested answer focus: explain separation of responsibilities, handling of concurrent or slow operations, and how API errors would be reported reliably.
What would you change when moving an LLM prototype to a cloud deployment?
What to coverEnd-to-end production deployment, Docker and cloud infrastructure are stated parts of the role. Suggested answer focus: discuss packaging, configuration, scaling and operational failure handling, while distinguishing a working prototype from a service that can be monitored and maintained.
Behavioural practice
Tell me about a time you took an AI or software solution from prototype to production.
What to coverThe posting calls for end-to-end ownership of AI solutions. Suggested preparation: use a real situation, your specific actions and the result; explain what changed between the prototype and deployment without claiming work you did not do.
Describe a time you investigated an unreliable workflow or production failure.
What to coverReliability, observability, logging and error handling are central to the stated work. Suggested preparation: give the situation, the evidence you used to isolate the problem, your actions and the observed result.
Tell me about a time evaluation results changed your technical approach.
What to coverThe role calls for assessing accuracy, reliability, latency and cost efficiency. Suggested preparation: describe the situation, the measurements and trade-off you considered, the decision you made and its result.
Describe a time you chose between a simpler implementation and a more flexible workflow.
What to coverThe posting spans scalable agentic workflows and production deployment. Suggested preparation: use a real situation, explain your decision criteria and actions, and state the result, including any limitation you accepted.