When I review candidates for AI work at SCULT INDIA or coach students before interviews, the questions that matter are the ones a tutorial cannot answer for you. Here are the ones I would ask, grouped, with what a strong answer contains. Prepare these and you will also have built the right things.

Retrieval

  1. How did you decide your chunk size? Strong: an evaluation set and recall numbers at three sizes.
  2. Why hybrid retrieval? Strong: an example query where vectors failed and keywords worked.
  3. What does your system do when nothing relevant is retrieved? Strong: a threshold and a refusal, tuned on unanswerable questions.
  4. How do you know your citations are correct? Strong: markers resolved to chunks, a faithfulness check.

Fine-tuning

  1. When would you fine-tune instead of using retrieval? Strong: behaviour versus knowledge, with an example of each.
  2. What did your loss curve look like, and when did you stop? Strong: eval loss, checkpoints, the best step.
  3. How did you prove the tuned model was better? Strong: base versus tuned on held-out data, same rubric.
  4. What was wrong with your dataset the first time? Everyone has an answer; the honest one is the strong one.

Reliability and cost

  1. What happens when the provider is down? Strong: fallback, retries with limits, a typed error to the client.
  2. How do you control cost per request? Strong: context caps, caching, cheaper routing for simple cases.
  3. What do you log for an LLM call? Strong: request id, model, tokens, latency, retrieval scores — never raw personal data.
  4. How would you test this system in CI? Strong: golden prompts, schema checks, an evaluation script with thresholds.

Product judgement

  1. Was AI the right tool for this feature? Strong: names the part done with plain code.
  2. What did you cut, and why? Strong: a deliberate scope decision.
  3. Who used it, and what did they say? Strong: named users and a change made because of them.
  4. Where does user data go? Strong: a one-sentence data flow, and what is never stored.

About you

  1. Show me something running. The best candidates have a URL ready.
  2. Walk me through the hardest bug. Reveals who built what.
  3. What would you build differently now? Reveals reflection.
  4. What are you learning this month? Reveals momentum.

Notice that every strong answer is evidence from something you built. That is the whole strategy: build the projects in the roadmap, and the interview prepares itself.

— Pranjul Rathour, GenAI Engineer from Kanpur, India. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at any campus: pranjulrathour41@gmail.com.