Skip to content
MVP operationalHuman-in-the-loop · Deterministic control plane

AI Operations Control Room

A Telegram-first control plane for researching, drafting, approving and publishing financial content. AI drafts; a state machine decides; a human approves.

Agents where useful. Deterministic software where mistakes have consequences.

01Context

A small editorial operation publishes daily macro and forex content to public channels. The admin works from a phone, between other tasks.

02Problem

Use LLMs to research and draft — without ever letting a model publish something on its own.

03Constraints

  • Publishing is irreversible and public
  • Admin interface must be Telegram
  • Must survive restarts without double-posting
  • Single VPS, Docker only

04My role

What I personally built

  • Router / worker / sender control-plane architecture
  • Approval state machine with opaque tokens
  • Outbox pattern — only the sender calls external APIs
  • n8n workflows for LLM drafting
  • One-command bring-up: Docker, migrations, workflow import, health checks
  • Handoff and operations documentation

What others owned

Independent project — no team. Everything here is mine.

05Architecture

Architecture explorer

The business flow — what happens, in plain words.

01 / 04

Admin asks — A command from Telegram.

06Key decisions

  1. Outbox pattern for every external side effect

    Because
    Exactly one component can publish; retries never double-post.
    Trade-off
    Extra table and a polling sender.

  2. Deterministic state machine; LLM only drafts

    Because
    Consequential actions stay auditable and predictable.
    Trade-off
    Less 'magic' conversational UX for the admin.

07System

Telegram updates hit a webhook router. A worker advances a PostgreSQL-backed state machine and calls n8n for drafts. Messages are written to an outbox; a dedicated sender is the only component allowed to talk to Telegram or Postiz. Approval requires an explicit callback carrying an opaque token.

08Challenges

  • No LLM auto-routing on free text — every transition is explicit
  • Idempotent delivery across restarts
  • Reproducible setup for a non-developer operator

09Outcome

MVP operational: draft macro post → approval callback → publish to a public Telegram channel, brought up with a single script.

10What I learned

The most important agent design decision is what the agent is not allowed to do.

11Stack

  • Node.js
  • Express
  • PostgreSQL
  • n8n
  • Telegram Bot API
  • Docker Compose
  • OpenClaw
  • Postiz
  • Apify
  • Webhooks

12Evidence