gora.
AI / Telegram assistants

GoraBot — multi-AI assistant on Telegram

12 models in one bot, 4-AI chains, AI debates with a judge. 56 users, 1,783 generations, 5.4M input tokens.

Region
@aiforum48_bot · global access
Role
Solo: architecture, development, ops
Period
2024–2026
Industry
AI / Telegram assistants

Numbers from the live /opt/multi_ai_bot/data/bot.db on publication date. User identifiers and message content are not published.

Summary. Below — why this architecture, what's inside, and why it stays small.

Why multi-AI in one chat. Every model has a profile: Claude Opus reasons deeply, GPT-5.2 Think is faster on technical tasks, Gemini holds a huge context, Grok offers a less polite but sometimes more honest view. Tab-hopping between chat.openai.com, claude.ai, gemini.google.com and x.ai loses context and time. GoraBot keeps all four providers in one Telegram chat under a single rate limit and billing logic.

Chains. Not "pick a model" — "run the task through all of them". Opus does the analysis → GPT-5.2 continues with the analysis on hand → Gemini adds context → Grok offers the contrarian view. Each model sees the previous results and avoids repeating them. The output is four perspectives on one task, glued into one thread.

Debates. Polemic between models. Set the topic, pick 2–4 participants, pick the style. AIs argue in rounds; in the end, GPT-as-judge delivers a verdict. Useful when you don't know which side to take: you walk away with three or four reasoned views and a third-party decision.

Projects. Context folders. Each project keeps its own conversation history and system prompt. Switch projects with one command — /project ge.hiregora.com and the bot is already in the landing-page context, remembering prior discussions. 14 live projects in the DB on publication date.

Architecture without overkill. python-telegram-bot 21 with aiogram-style handlers, SQLite for everything, one client module per provider, systemd unit, bot.py.backup for rollback. No Kubernetes, no Docker, no Redis — this is a bot for 56 users, and every layer of complexity would have cost a week of work instead of two days.

What's inside

  • Twelve models in one chat: Claude Opus 4.5 / Sonnet 4.5 / Haiku 4.5, GPT-5.2 Think / Fast / Mini, Gemini 2.5 + 3 Pro/Flash, Grok 3 / Mini.
  • Chains: four AIs process the same task one after another (Opus → GPT-5.2 → Gemini → Grok), each sees the previous results. Four modes — Analysis, Code Review, Creative, Summary.
  • Debates: 2–4 AIs argue on a given topic. Three styles — Balanced (constructive), Aggressive (no holds barred), Socratic (truth through questions). GPT judge delivers the verdict.
  • Project contexts: 14 active projects in the DB, each with its own conversation history. Switch context with a single command.
  • Tariffs: Free (10/day), Basic (50/day), Pro (unlimited), Admin. Limits enforced at the DB level, request_logs for analytics.
  • Real numbers from the production SQLite: 56 users, 1,783 generations, 5.4M input + 1.06M output tokens, $43.3 in API costs.
  • Voice replies via TTS — long summaries and narratives are spoken back directly in Telegram.
  • Top models by usage: GPT Thinking 605, Claude Sonnet 492, Claude Opus 311, Grok 146, Gemini Pro 124 — real weights from the usage table.
  • Inline buttons swap the model mid-thread: cycle through 12 engines in a single conversation without losing context.
  • Backup copies of every client module live alongside the originals (claude.py.backup, gpt.py.backup, etc.) for fast rollback when an API provider ships a breaking change.

Stack

  • python-telegram-bot 21
  • anthropic SDK
  • openai SDK
  • google-genai
  • xAI Grok API
  • SQLite + aiosqlite
  • FastAPI (webapp shell)
  • systemd
  • Ubuntu 22.04 + VPS
Three narratives

One project — three angles

    GoraBot — multi-AI assistant on Telegram · hiregora.com