I just finished reading the first chapter of "Introduction to AI Engineering" by O'Reilly, and I decided to stop just reading and actually build something. I wanted to take the concepts of capabilities and control and implement them into a system that actually respects me.
This isn't just another chatbot wrapper. It is an experiment in digital sovereignty which I call rei.
The core idea I'm exploring here is what happens when you don't trust a single model blindly. We know LLMs hallucinate and can be weirdly manipulative if trained that way. So instead of one model, I built a Council.
I have two distinct AIs running in parallel:
- Groq (LPU) for speed and reasoning
- Ollama (running locally) for privacy checks
When I type a command, it doesn't just go to a server and back.
First, the Council convenes. Both models come up with an answer independently. Then I have an Auditor script that scans their responses for any "soft coercion" or manipulative language. If it smells like the AI is trying to gaslight me or nudge me into something I didn't ask for, the Auditor flags it.
Finally, there's the Memory. This is the coolest part. I hooked up a local vector database (ChromaDB) that remembers my boundaries. If I reject an answer, rei remembers that specific context. It won't try to cross that line again. It's a system that learns to respect my "No."
It's a Python CLI because I wanted to keep it clean and focused on the engineering, not the UI.
You'll need ollama running with llama3.2:1b (or whatever local model you prefer).
pip install -r requirements.txt
python src/app.pyThat's it. No fancy dashboards, just a terminal that listens.