A rules-first moderation guard for Telegram groups. It acts only where a group admin has added it, and every call it makes comes from a rule a person can read.
Security tooling that decides on its own is hard to trust. SAT Digital keeps every decision inspectable and every action inside a group someone chose to protect.
Telegram sends each group update to a serverless webhook. The webhook checks a shared secret first, then applies the same inspectable rules the local build uses, and deletes only what those rules name: configured scam, phishing, impersonation, spam and sensitive-content patterns.
It acts only in a group where a human has added it and made it an admin with permission to delete messages. The rules catch what they are written to catch; a miss is a rule that has not been written yet, not a model's judgement call.
The repository holds two guard paths. One is deployed, and nothing new ships until the current paths are verified.
The Telegram group guard, running as a serverless webhook with rules-only detection and moderation actions. It answers on Telegram today, with no pilot, users or protection outcome behind it.
A server-side check that returns a bounded pass or block for submitted text, with a review queue and rate limiting. It stays local, and an owner must authorize any site it is ever pointed at.
A fine-tuned classifier needs reviewed, lawful, representative signal the project does not have yet. New guard ideas wait too, until each has one authorized surface, one safe action and one accountable owner.
The next finish line is repeatable verification and a narrow, authorized pilot — not more guards or a broad launch.
Before a pilot, the team must verify both current guard paths, document what they detect and may change, obtain authority for every target, protect logs and user data, rehearse monitoring and rollback, and define escalation when the tool is uncertain.
SAT Digital must not promise complete protection, silently act outside an approved surface, or treat a model decision as final authority. Two client website pilots do not weaken these security gates or move this work ahead of the active completion lane.
A working security product needs accountable human review. These are the hands responsible for its current direction.
The natural fit. Pagna owns security and trust across everything Chnai Lab makes, so the rules that guard other people's groups are his to design — what they watch, what they catch, and where the line is drawn.
Hongleng supports the detection layer. Rules remain primary; model-assisted classification is a backstop that must be reviewed against false positives, Khmer-language cases, privacy limits, and the owner's policy.