314,552 interview questions from 6,000+ companies.
Explain how to evaluate a generative model using offline and online methods, with attention to hallucination, product metrics, and experiment design.
Design a grounded document Q&A system and explain how vector search improves retrieval quality, latency, and hallucination control in RAG.
Explain LLM hallucination and give three practical ways to reduce it using grounding, prompting, and evaluation.
Explain a practical approach to fine-tuning an LLM for a specific task, including data, evaluation, and hallucination risks.
Design a grounded multi-agent assistant that plans, retrieves, and synthesizes answers under strict latency, cost, and hallucination limits.
Compare RAG and fine-tuning, and decide when each is the better fit for an LLM product.
Discuss how you designed an LLM system for a business use case, including evaluation, hallucination control, and cost latency tradeoffs.
Discuss the main ethical risks in deploying generative AI, including hallucination, misuse, privacy, and governance.
Design a production agent platform that coordinates models, tools, and data sources under strict latency, cost, and safety limits.
Design and evaluate a RAG assistant over internal policy and delivery docs with strict latency, cost, and hallucination limits.
Reduce hallucinations in a RAG system even when retrieval is already correct, using grounding, verification, and evaluation.
Explain prompt engineering and RAG, how they differ, and when each is useful for improving LLM answer quality.
Diagnose why a customer-facing LLM assistant is underperforming, using eval-first debugging across retrieval, prompting, safety, latency, and cost.
Design a safe LLM workflow that explains prompt injection to technical customers without hallucinating or overstating security guarantees.
Explain what RAG is and how it reduces stale, ungrounded answers in enterprise AI systems.
Talk through a real generative AI project, focusing on architecture, evaluation, hallucination risk, and how you handled safety issues in practice.
Design a grounded LLM assistant that cuts unsupported claims below 2% while meeting strict latency, cost, and safety limits.
Explain how self-attention works and why it is central to transformer-based LLMs.
Design state management for a multi-turn agent with context window limits, durable memory, and low-loss summarization of user preferences.
Explain how to manage memory, summarization, retrieval, and safety in a long-running LLM agent when context exceeds the model window.
765 total questions
