What's a practical guardrail stack for a consumer-facing AI product?

We've moved past "just add a system prompt" for safety. We need a multi-layer approach: input classification, output filtering, jailbreak detection, and fallback routing. We've looked at Guardrails AI, Llama Guard, and building our own classifiers. The challenge is keeping latency under control while catching edge cases that slip through vendor solutions. What does a production guardrail stack look like for consumer scale?