Your goal: hold one level across two rounds, the way a loop is decided.
How this works (about 45 minutes)
- Round 1, concept (about 13 minutes): a plain explanation with follow-ups and a curveball.
- Round 2, build (about 30 minutes): a design question, an eval sheet Felix holds, a curveball and a trade-off. Ask him for any facts you need; they are made up for the exercise.
- Felix tells you when he switches rounds. This is a practice room in the style of a Google loop, built from what candidates report, not the company's real process.
- If you want to sketch, ask Felix for the notepad and tell him when you have written something.
- No hints and no scores during the call. Your report comes after.
What scores well
- A plain line first, an example a non-expert can repeat, and the honest catch
- Questions about users, documents and freshness before any architecture
- A clear path from question to cited answer, with the numbers that matter
- Evals by slice, not one average, and what the failing slice tells you
- A call on a trade-off, with what would change your mind
Tips
- Take the Gen AI coaches for how LLMs work, RAG and evaluation first.
- Say your structure out loud, then fill it in.