Skip to content

Latest commit

 

History

History
39 lines (28 loc) · 1.05 KB

File metadata and controls

39 lines (28 loc) · 1.05 KB

Interview Questions

Beginner

  1. What is the difference between training and inference?
  2. What is a token?
  3. Why does a longer prompt usually cost more?
  4. What is time to first token?
  5. Why can an LLM return different answers to the same prompt?

Intermediate

  1. What are prefill and decode?
  2. Why does KV cache use GPU memory?
  3. What metrics would you track for an LLM service?
  4. How would you benchmark a model server?
  5. What are the risks of quantization?

Production

  1. How would you debug high p99 latency in an inference service?
  2. What autoscaling signals are useful for LLM serving?
  3. How would you roll out a new model safely?
  4. How would you reduce inference cost without hurting quality?
  5. What should be logged, and what should not be logged, for privacy?

Practical Exercise

Design a chat service that uses an LLM to answer questions from company documents.

Explain:

  • How documents are retrieved.
  • How prompts are built.
  • Which metrics are collected.
  • How output is validated.
  • How the system scales.
  • How cost is controlled.