Crime Sentinel

The most dangerous True Crime Community spaces aren't public. They sit behind admission questions ("are you a journalist?") and invite-only chats, where keyword filters and public scraping can't reach. Crime Sentinel sends in a governed AI research persona, powered by Meta Muse, to find communities that glorify violence before they act: 🕵️ A Muse-written persona answers a closed group's admission questions and is later invited into a secret chat 🛡️ An independent Safety Governor (hard rules plus a separate Muse judge) checks every message. The persona never praises, encourages, or commits to anything 👁️ Muse vision reads posts and images, then one long-context call rates the whole community 📨 Muse drafts an evidence-backed report to Meta Integrity, and a human approves it before it is sent 🔍 Every observed post flows into a human review console with explainable classification, a GraphRAG network, trends and evaluation On our test set, Muse flagged no harmless posts as harmful. A keyword filter flagged 45%. It assesses content and communities, never people: no identity resolution, no individual threat scores, and no automated reports to police or schools. 🧰 Built with: Meta Muse (muse-spark, muse-image), FastAPI, React, PDQ perceptual hashing, GraphRAG (Louvain), multilingual embeddings, Cytoscape