← Back to Projects

Hermes: Telegram Business System

A Telegram system where AI prepares drafts with channel context, but cannot quietly publish or reply.

AI drafts stay useful only when risky sends remain reviewed.

WHERE IT BREAKS

AI drafts, audience trust, and delivery collapse into one bot action.

WHAT CHANGED

Drafts, channel memory, and sends become separate visible steps.

WHAT YOU CAN SEE

Draft status, channel context, review rules, and delivery limits are visible.

draft before send
Risk
channel memory
Context
reviewed send
Action

Where It Was Risky

Businesses relying on Telegram for community engagement face fundamental scalability constraints:

  • Manual Moderation Bottleneck: Human operators cannot maintain 24/7 engagement across multiple channels.
  • Platform Policy and Reputation Risk: Automation around public communities must respect platform limits, protect audience trust, and keep brand review boundaries explicit.
  • Context-Blind Responses: Simple keyword-triggered bots produce generic replies that damage brand perception.
  • Role Management: Business channel work needs visible status, explicit permissions, and conservative delivery controls.

What Changed in the System

CONTENT FLOW
  Channel Monitor → Context Extractor → Draft Generator (GPT route)
          ↓
  REVIEW LANE
  Platform Rules | Review Queue | Action Log | Delivery Settings
          ↓
  SEND PATH
  Context Tracking | Approval Check | Quality Check | Incident Notes
          ↓
  TELEGRAM API (MTProto via Grammers)

What Has to Work

  • Channel Monitor: Real-time event stream from target Telegram channels via MTProto.
  • Context Extractor: Semantic analysis of post content, thread history, and channel tone.
  • Draft Generator: Context-aware draft synthesis with persona-specific voice, tone matching, and review-ready output.
  • Review Rules: Publishing limits, review steps, action logs, and delivery-policy handling.
  • Send Review: Structured status for context tracking, approval checks, quality review, and incident notes.

What Became Visible

draft before send
Risk
channel memory
Context
reviewed send
Action
checked GPT route
AI Route
draft + send step
Work Status
review limits
Audience Trust
Rust + Tokio
Runtime

Decisions That Remove Risk

Problem
Community operations need several permissioned working identities without losing auditability or confusing production ownership.
Solution
External context is turned into clear draft records with an owner, review status, and conservative delivery settings.
Alternative Rejected
Unstructured handling — makes permissions, support, and incident analysis harder.
Problem
Unpaced publishing can overwhelm audiences, reduce trust, and create avoidable platform friction.
Solution
Conservative delivery settings: slower pacing, review for risky actions, and explicit approval before public posting.

Tools Behind It

Backend
Rust, Tokio, Grammers (MTProto), Axum
AI
Checked GPT route via Antigravity Gateway
Data
PostgreSQL (event sourcing)
Review
Platform rules, review queues, action log

Why It Matters

Context-to-Action Orchestration
External channel signals become reviewable draft work, not uncontrolled bot behavior.
Checked GPT in a Live Channel
Model output is useful only because routing, source context, review status, and delivery policy stay explicit.
Telegram as a Business Channel
The channel is treated as a business environment with identities, permissions, pacing, and review needs.
Approval Stays Visible
Drafts, risky actions, delivery readiness, and incident notes remain separate enough for humans to keep authority.
Audience-Safe Delivery Loop
The automation reduces operator load while protecting audience trust and platform boundaries.
> Private repository. Available for code review on request.

Ready to build something like this?

Start a Project