Design an LLM content safety service

An LLM application checks every user prompt, retrieved context and model response with a safety service before showing anything to the user. Design that service. It must catch plain policy violations and LLM-specific attacks such as jailbreaks, obfuscated text and harmful intent that builds up over several turns, and answer allow, block, redact or send to human review.

Cover:

• The low-latency decision path and how multi-turn context reaches it.
• Versioned policies and thresholds that can be rolled out and rolled back, with every decision recording which versions made it.
• A reviewer queue for uncertain cases whose verdicts feed back into the detectors.
• Audit and quality metrics without exposing raw conversation text widely.