Elsa – Global Banhammer

Elsa protects Telegram communities from a single problem: child sexual abuse material. It analyses message text and matches images against material already flagged, and when something looks suspicious it flags the sender for review — the bot can mute for an hour to stop the harm spreading, but removing an account is always a moderator's call, and that decision applies across every group in the network at once.

Underneath sits a purpose-built classification model, trained on real messages that a closed group of reviewers labels by hand against written criteria. The interesting part is how the right label gets decided when two people disagree: instead of counting votes, a statistical estimator works out each reviewer's reliability from their own track record and weights their vote accordingly. A set of cases with known answers keeps the calculation anchored, and a label is only accepted once the system is confident; anything still in doubt goes to arbitration.