Production LLM: Architecture, Domain Tuning and Hallucinations

HardLLMs & GenAI

Reported in interviews at NVIDIA

Problem statement

How would you adapt an assistant to answer questions about a customer's insurance policy? You have a pretrained, instruction-tuned decoder-only transformer, domain documents and expert-written question-answer pairs. Answers must use the customer's own policy passages retrieved into the prompt.

Explain the architecture, domain tuning and how to keep answers grounded.

Show hint

For each part, start from the specific failure the component prevents: what goes wrong without attention, without replay, without verification?

Relevant roles: ML Engineer, LLM Engineer, AI Engineer and Gen AI Engineer.

Related free problems