You already know: the LLM gateway (where a router lives) and evaluation with model judges from earlier cases. This case asks a deceptively simple question: which model should answer this request? The hard parts are predicting difficulty before seeing an answer, measuring quality on answers you did not generate, and staying correct as models and traffic change.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "classify difficulty with keywords and route". A staff candidate asks: what does 'well' mean, and how will we measure it, what are the costs of each kind of mistake (sending a hard question to a small model versus an easy one to a large model), how much latency may the router add, and who is allowed to override it.
Assumptions you state: