Skill 05
Local AI integration
AI where the data never leaves the building – and that may only do what it should.
Language models now run well on a single graphics card. That makes it possible to build assistants that summarise emails, search documents or answer voice messages without a single byte ending up with a US provider.
The hard part is not the model but control: which tools may the agent use, who may it send to, and what happens when a message it reads contains instructions? That is where I put the focus.
What I do
- Setting up local language models and connecting them to your data
- Assistants for Telegram, email and internal tools
- Voice in and out: Whisper and Piper running locally
- Security concepts against prompt injection and data leakage
- Making documents and scans searchable with OCR
Tech stack
| Layer | Technology | Why |
|---|---|---|
| Models | Ollama (e.g. Qwen3) | swappable, local, no API costs |
| Voice | Whisper (STT), Piper (TTS) | understand and answer voice messages |
| Documents | Tesseract, pypdf | PDFs and scans as text |
| Control | custom policy engine in Python | deny-by-default, approval per tool |
| Operations | Docker, optional GPU | one stack running on your own hardware |
Related projects
Planning something similar?
Tell me by email what you have in mind. The first conversation is free.