In brief: An assistant in Python that decides for itself which language model gets each task: GPT, Claude or Gemini. It reads documents, walks into servers over SSH, edits spreadsheets, drives a browser and runs shell commands. It is controlled from a web interface or straight from Telegram, and the whole system is deployed on the client’s own infrastructure, so the data goes nowhere.
Client: private individual · Platform: Telegram and a web interface · Category: AI assistant
What matters most in this project?
- More than 100 operations in one assistant, from text generation to server monitoring.
- Three models instead of one: GPT, Claude and Gemini, picked for the task at hand.
- Local deployment on the client’s infrastructure: data and processes stay under their control.
- Two control channels: a web interface for complex work and Telegram for anywhere else.
- Its own character and logic, tuned to the client’s workflows and way of speaking.
- Access to real working tools: SSH, the file system, Google Sheets, a browser, PDF processing.
What is the project built with?
- Language:
- Python.
- Models:
- GPT, Claude and Gemini with dynamic selection based on the task.
- Deployment:
- local, on the client’s infrastructure.
- Interfaces:
- a web interface and a Telegram integration.
- Integrations:
- external APIs, SSH, Google Sheets, a web browser, the file system.
Different models are strong at different things: one holds a long context better, another is more precise with code, a third is faster and cheaper on simple queries. So the assistant is not tied to one vendor and picks the executor per task.
How did it start?
An ordinary chat with a neural network solves half the problem. It suggests the command but does not run it; it writes the text but does not put it in the right file; it explains how to check a server but never goes there.
The client needed a doer rather than a conversationalist: one message in Telegram should end in a real action on their machines and in their documents.
The second requirement was tougher than the first. Working files, server credentials and internal data all pass through the assistant, so a cloud build was out of the question. The system was deployed locally, on the client’s infrastructure, where they control both the processes and the storage.
What was built?
🧠 Intelligence and language models
- GPT, Claude and Gemini combined with dynamic model selection
- The assistant’s own logic and character, tuned to one specific person
- Advanced search, content generation, data analysis
- Image generation
⚙️ Automation and system operations
- Shell command execution
- File management
- Web browser automation
- PDF processing
- Server monitoring over SSH
- Working with Google Sheets
💬 Access channels
- A full web interface for complex operations
- A Telegram integration: control from anywhere in the world
🛡 Security
- Local deployment on the client’s infrastructure
- Full control over processes and data
What is the result?
| Metric | Value |
|---|---|
| Operations in the assistant | more than 100 |
| Language models | 3, chosen per task |
| Control channels | 2: web and Telegram |
| Where it runs | the client’s infrastructure |
| Data in external services | none stored |
| Working integrations | SSH, files, Google Sheets, browser, PDF |