Roblox Sentinel

Sentinel is the system Roblox built to catch grooming and child endangerment in chat, released as open source. It doesn't look for individual bad words. It embeds a user's messages over time, compares those patterns against what known offenders and known safe users look like, and flags accounts whose behavior drifts toward the risky cluster. That's what makes it useful for grooming, which rarely trips a single message classifier.

You'll need to adapt it to your own data and reference sets; the code and approach are open, the Roblox training data isn't. It's best thought of as a reference architecture and starting point for a behavioral detection system, not a plug in filter.