"To prepare for this, I built a working on-device RAG tutor. Retrieval runs locally (BM25 over a curriculum I authored), and generation runs through a local Claude-Code bridge on my own machine on my subscription — so no sensitive data ever leaves the device and there are no per-token API costs. It does cited Q&A, mock-interview grading, and flashcards. I used it to study, and it's a direct, small-scale version of the privacy-first architecture this role needs: keep the sensitive data on-device, send only the minimum necessary for generation, cite everything."
This reframes the interview: you didn't just study RAG, you shipped one.