You already know: the LLM gateway (routing, quotas, logging), the guardrail gateway and evaluation gates. This case adds privacy: making sure sensitive data reaches only approved models, in approved regions, with as little retained as possible. The design questions: how to classify data, where each class may go, how to hide personal data without breaking tasks, what to log and for how long, and what to do when detection misses.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud. This is a design discussion, not legal advice: data rules come from your legal and privacy teams.
1. The prompt, and the questions that change the design
A mid-level answer: "detect personal data with a model and redact it before calling the provider". A staff candidate notes that detection is probabilistic: even at 99% recall, thousands of prompts a day slip through. So the design needs deterministic controls behind it: classify data by where it came from, route each class only to destinations allowed to receive it, and rely on contracts (no retention, region) for what detection misses. They also note that usability is a privacy control: if the gateway is painful, people paste data into unmanaged tools.
Assumptions you state: