Automation · Multi-agent · Demo
AI Marketing Team
A multi-agent Telegram assistant for a coffee shop that writes posts, manages the content calendar and analyses performance, by text or voice. Built in n8n as a demo with sample data.
Demo project with sample data, built around a fictional café, "Ben Al-Sabah".
Problem
Small café owners know they should post every week, but they don't have time to write captions, keep a posting calendar, or check which posts actually brought customers in. Generic AI tools don't help much: they don't know the café's menu, prices, offers or tone of voice, and they make things up.
Solution
A marketing team in the owner's pocket. They send one Telegram message, typed or spoken. An Orchestrator agent understands the request, delegates to Content, Calendar and Analyst agents, reviews their results and replies with one final answer.
Architecture
Live replay
Scheduling a post, with the owner's approval
What it does
- Writes and edits posts for Instagram, Facebook, X and WhatsApp in the brand's voice
- Prices and offers come only from a RAG knowledge base. Missing information is reported, not invented
- Shows the calendar, suggests times, and rejects past dates and taken slots in code
- Human approval gate: nothing is added to the calendar until the owner approves, enforced by the workflow itself
- Answers performance questions from the sheet, using a calculator for every figure and citing the posts used
- Voice in, voice out, in Arabic (Levantine) or English, detected automatically
- Owner-only allowlist, friendly error messages and failure alerts
Results
- Content, calendar and analyst requests routed to the right agent in every test run.
- Prices and offer dates in generated posts matched the knowledge base, and an offer that had not started yet was correctly left out.
- Calendar conflicts were detected, and posts were saved only after approval.
- Performance answers matched hand calculations, such as an engagement rate of 19.76% for one post.
- Unknown information, such as a delivery service, was reported as missing instead of invented.
- Asked to announce a discount code that wasn't in the offers file, it stopped and asked the owner to confirm the details instead of publishing it.
From testing with sample data.
Challenges and fixes
The model sometimes answered English questions in Arabic, because the knowledge base is in Arabic.
The reply language is computed in code and passed to every agent as a required instruction.
One long RAG query returned platform rules instead of the product price.
The Content Agent makes short, separate searches per topic, and searches twice before saying information is missing.
The free Gemini tier allows 5 requests per minute, and one multi-agent request needs about 5.
Chat models moved to OpenAI, with Gemini kept for embeddings so the vector index didn't need rebuilding.
The first approval flow depended on the model remembering that the owner said yes.
Replaced with a Telegram approval node that pauses the workflow until the owner approves.
Telegram rejected some replies because of underscores in hashtags.
Replies are sent as escaped HTML, with Markdown characters escaped where HTML isn't available.
In action
(opens the full image)
(opens the full image)
(opens the full image)
(opens the full image)Stack
- n8n
- OpenAI (gpt-5-mini)
- Pinecone
- Gemini Embeddings
- Google Sheets
- ElevenLabs
- Telegram Bot API
My role
AI automation design and build, solo.
What I'd add next
- Approve buttons directly inside the chat
- Edit and cancel existing calendar entries
- Automatic publishing to Instagram and Facebook
- A small performance dashboard