Potential incitement, triaged on ared-amber-green scale before it spreads.

Online incitement is hard to catch early: risky content surfaces across social posts, articles, political commentary and public complaints. In election periods and moments of public tension, harmful narratives spread fast and contribute to real-world risk.

MMA needed a repeatable, evidence-led way to identify potential harms before they become harder to manage — without treating every piece of content as equally urgent.

A custom AI classifier with sentiment analysis, AWS-based model training and a Next.js dashboard, assessing content on a Red-Amber-Green risk scale: Red for highest risk, Amber for moderate, Green for low or none.

Legal professionals guided data collection and annotation, linguists refined the language features, and data scientists trained the model through a supervised MLOps process, with 90% of the data for training and 10% for testing.

Earlier triage and better prioritisation: researchers and stakeholders work higher-risk content first, using the RAG classification as a practical risk lens.

The manual review burden dropped, and the dashboard made digital harms monitoring repeatable and trend-aware over time.

Human judgment stays in the loop for interpretation, legal nuance and responsible use — a defensible technical base for public-interest research and policy engagement.