I'm a SW/ML engineer that spends a tad too much time on Twitch.. I've been eyeing the moderation tooling used for livestreaming and I can't help but notice how primitive it still is - keyword lists, regex, basic shimming, etc. So, I thought I could do better.

Behold ModCheck - a context-aware moderation tool. It understands what's happening in the stream, who the streamer is, what's currently happening, the mood of chat, what the streamer is saying, etc.. It uses all of that to make moderation decisions.

Instead of keyword lists or regex, you write moderation rules in plain English. I call them Judges. They run in parallel for each message. A few examples: 1. Boundary Judge - "Flag viewers who comment on my appearance, body, or clothing choices." Catches stuff like "you'd get way more viewers if you showed more skin js" 2. Backseat Judge - "Flag unsolicited gameplay advice and strategy tips. Allow tips if I've asked chat for help."

The goal is to make this fully open source and have a paid tier for those who don't want to self-host.

NOTE: the product is not yet released. I'm looking for early adopters to test it out and help shape the direction. Also, if anyone is interested in building this out with me, please let me know!

--

For the interested, Judges are based on 1) scemantic similarity over a space of previously flagged examples having the stream context embeeded into the search and 2) finetuned small models as classifiers. The combination of the two allows to keep latency low-ish (under 500ms) and to use less GPU. Also 1) grows with your usage. The more you use a judge, the more accurate and fast it becomes.