Design a RAG bot over 10M docs, an AI coding agent, an eval platform, a voice assistant, an agentic research system.
An eight-stage pipeline you drive in any AI design interview — where the LLM call is only ~20% of the system and tokens, freshness, and hallucination are the real constraints.
The canonical AI system-design question, fully worked — from ingestion and chunking to vector-store choice, hybrid retrieval, freshness, access control, and the cost model that decides which one wins.
A coding agent is a permission-gated loop wrapped around a sandbox, where the tests are the reward signal and the hardest engineering is keeping a dangerous tool surface safe at scale.
A version-controlled judge harness that turns "the model feels worse this week" into a red CI check with a diff you can point at.
A voice agent is a latency machine wearing an LLM costume — you win by overlapping STT, generation, and TTS inside an 800ms budget while still answering the phone when the user interrupts.