I ended up building ELI around a pretty simple idea: don't trust the LLM to describe system state.
So i am looking for some feedback if posible, hopefully this project works for you and helps you out in daily life! So poke holes, ask questions, and get involved if you want to participate!
At this point it can:
Desktop automation
Read your screen, PDFs and images
Local voice in/out with your own wake word
Layered memory (SQLite + FAISS + knowledge graph)
Generate document-grounded reports with citations
Sandboxed code execution
Local FastAPI/PWA interface you can open from anything on your LAN
MQTT smart-home integration
Runs locally and offline by default
It's mostly just been me over the guts of a year, so it's definitely not finished and I'm under no illusion there's plenty I've missed. That's kind of why I'm posting it here, just to get some honest feedback and to see if this project helps people.
I'd really appreciate people pulling the architecture apart, pointing out bad ideas, suggesting improvements, or just generally trying to break it. I'm especially interested in how other people separate deterministic system state from probabilistic LLM reasoning (if you do at all). One last thing! Check out the readme's, and the documents in /blueprints for a much clearer idea!