Content moderation without the burnout
A streaming platform with a growing UGC layer needed to scale moderation without scaling headcount or risk.
A streaming platform with a growing UGC layer needed to scale moderation without scaling headcount or risk.
User-generated content was growing fast. Human moderators were drowning in queues; turnover was high and consistency was hard. The platform had community guidelines but no unified way to triage content by risk, auto-flag clear violations, and surface edge cases for human review with context. They needed to protect the community and their brand without turning moderation into an unsustainable cost.
We built a moderation pipeline that ingests text, thumbnails, and metadata. It runs policy-aligned classifiers for harm categories (hate, abuse, spam, NSFW, etc.) and assigns a risk score and suggested action. High-confidence violations can be auto-hidden or escalated; borderline content goes to a review queue with the model’s reasoning and relevant policy snippets so moderators can decide quickly. We tuned the models on their guidelines and a labeled set of their data so false positives stayed low and moderators could trust the triage. The system is designed to retrain as policy and language evolve.
Moderation throughput went up without adding proportional headcount. Moderators spend more time on nuanced cases and less on obvious violations. Appeal rates and policy breaches both improved. The platform is now extending the same stack to live and near-live content and to new regions and languages.