Member-only story
I watched a friend walk into a senior AI engineer loop last month with a portfolio full of solid RAG projects and a Medium-article-level understanding of agents. He drew a clean retrieval pipeline on the whiteboard, explained cosine similarity without stumbling, and felt good about it. Then the interviewer asked what happens when the retriever pulls back a document that contradicts what the user actually meant. He said he’d tune the prompt. He didn’t get the offer.
That question wasn’t a trick. It’s the new baseline. Reports from candidates interviewing at Anthropic, Scale AI, Sierra, xAI, and Glean through 2026 show the same pattern: RAG fundamentals are no longer the bar, they’re the entry fee. The real interview starts one level up, at system design questions that assume you already know what a vector database is and want to know what you do when the system you built breaks in a way a tutorial never covers.
This piece walks through seven questions that show up across those loops, again and again, with the kind of model answer that actually survives a follow-up question. Not definitions. Architecture.