InterviewCompany
LLM systems, evaluation, safety, RAG, and product-quality AI engineering questions.
Use task success, factuality, refusal quality, latency, cost, and human review.
LLM evaluation should combine golden datasets, rubric-based judging, production feedback, safety checks, and regression tests for known failure modes.
Ingest docs, chunk, embed, retrieve, rerank, generate with citations, and evaluate.
Mention permissions, freshness, chunk strategy, hybrid search, reranking, context packing, answer grounding, and feedback loops.
Prompting changes instructions at runtime; fine-tuning changes model behavior through training examples.
Prompting is faster and flexible. Fine-tuning helps style, format consistency, and repeated task behavior, but needs data quality and evaluation.
Optimize model choice, prompt length, retrieval, streaming, caching, and parallel work.
Latency work includes measuring time to first token, token generation rate, retrieval overhead, network cost, and fallback paths.