// ai-ml-interview-questions
ML systems, model design, and the applied-AI loop: RAG, agents, evals, and infrastructure.
8 min read The Positron AI interview, reconstructed from its open roles and memory-first chip design: what each engineering loop probes and how to read the comp. Read article
8 min read What Chai Discovery screens for across its ML research and research-engineering loops, with a realistic round structure, sample questions, and how to read the comp bands. Read article
7 min read A wrong RAG answer is either a retrieval failure or a generation failure, and the interview is mostly about proving you can tell which. Read article
7 min read The questions AI companies ask about getting reliable JSON, valid tool calls, and trustworthy output out of a language model. Read article
7 min read How speculative decoding speeds up token generation without changing model output, and the questions inference interviews use to test whether you understand it. Read article
7 min read How mixture-of-experts models split parameter count from compute per token, and the routing, memory, and serving tradeoffs interviewers test. Read article
6 min read The weight, activation, and KV-cache tradeoffs interviewers probe when they ask how you would shrink a model to fit on one GPU. Read article
7 min read The ML platform interview question that trips up strong modelers is how you build training data without leaking the future, and how you serve the same feature online. Read article
7 min read MCP interviews at AI-tooling teams spend less time on wiring a tool server and more on the trust boundaries around tool descriptions, transports, and per-user auth. Read article
7 min read The memory math, the parallelism strategies, and the failure questions AI-infra and frontier-lab interviews use to tell whether you have actually trained a large model. Read article
7 min read A walkthrough of the prompt injection questions AI and security interviews ask, why the attack has no clean fix, and the defense-in-depth answers that land. Read article
6 min read A concrete look at what fine-tuning interviews probe about LoRA, QLoRA, and the cases where fine-tuning is the wrong tool. Read article
8 min read Prompt injection sits at the top of the OWASP LLM list because there’s no complete fix, and interviews for AI-security roles are built around what you actually do about that. Read article
7 min read A grounded look at Hebbia’s engineering interview loop, the retrieval and applied-ML questions they ask, and how to prepare for a company that thinks standard RAG is a dead end. Read article
6 min read The quantization, pruning, and runtime questions that come up when the job is shipping a model to hardware instead of training one. Read article
What's missing, wrong, or worth adding? One line helps.
Sent — thank you. We read every one.
