20 questions AI teams actually ask

From real interviews for AI Engineer, AI Platform, LLMOps and MLOps roles. Each one shows what the interviewer is really testing. Write your own answer first, then connect it to something you have built, because that is what separates a good answer from a memorised one.

1. 1How do you monitor an ML model in production?

   What they are testing

   Whether you watch the model and its data, not just the server. Uptime alone is not an answer.

2. 2Explain data drift versus concept drift, and how you detect each.

   What they are testing

   That you know the inputs can change without the answers changing, and the reverse, and that each needs a different signal.

3. 3A model performs well offline but poorly in production. How do you debug it?

   What they are testing

   A calm, ordered way of narrowing down a problem across data, features and serving, instead of guessing.

4. 4Canary, shadow and blue-green deployment. When do you use each?

   What they are testing

   Whether you match the rollout to the risk, and know which one lets you test a model on real traffic without users seeing it.

5. 5How do you guarantee reproducibility in an ML system?

   What they are testing

   That you think about code, data, environment and randomness together, not just pinning library versions.

6. 6Explain precision, recall and F1. When do you optimise for each?

   What they are testing

   Whether you tie a metric to the cost of a wrong answer for the business, not just recite the formulas.

7. 7Fine-tuning versus RAG. Which and when?

   What they are testing

   That you can tell knowledge problems from behaviour problems, and you weigh cost and how often the data changes.

8. 8What is the KV cache and why does it matter?

   What they are testing

   Whether you understand what actually uses GPU memory during inference, which drives most serving decisions.

9. 9What are the main failure modes of LLMs in production?

   What they are testing

   Breadth. They want to hear about more than hallucination, and how you would catch each one before a user does.

10. 10Walk through designing a RAG pipeline.

    What they are testing

    A clear structure from ingestion to answer, and whether you mention how you would measure retrieval quality.

11. 11How do you prevent hallucinations in a RAG system?

    What they are testing

    That you treat this as a layered problem across retrieval, prompting and checking, and admit you can reduce it but not remove it.

12. 12What is prompt injection and how do you defend against it?

    What they are testing

    Whether you think like an attacker, and design limits on what the model can do instead of trusting a clever prompt.

13. 13How much GPU memory does a 70B parameter model need, and how do you size a serving node?

    What they are testing

    Back of the envelope maths you can do out loud, and that you remember memory is needed for more than the weights.

14. 14When does self-hosting beat a managed API?

    What they are testing

    A real cost and risk comparison, including the people needed to run it, not a preference.

15. 15How do you structure an evaluation suite and where do the gates sit?

    What they are testing

    Whether you treat evals like tests in CI, with clear points where a bad change is stopped before release.

16. 16How do you route between models by complexity and cost?

    What they are testing

    That you can save money without hurting quality, and that you would measure both before and after.

17. 17Walk through how a container gets access to a GPU in Kubernetes.

    What they are testing

    Depth on the layers between the hardware and the pod. This separates people who used GPUs from people who ran them.

18. 18Time-slicing versus hardware partitioning for GPU sharing. What is the difference?

    What they are testing

    Whether you know the trade between using every bit of an expensive GPU and keeping teams isolated from each other.

19. 19How do you manage schema changes in a production pipeline?

    What they are testing

    That you plan for change before it breaks things downstream, including the models and dashboards that read the data.

20. 20How do you make a pipeline idempotent, and why does it matter?

    What they are testing

    Whether you design for reruns and failures as normal events, so a retry never creates duplicate or broken data.

The model answers, mock interviews and system design frameworks are inside my community.