The workflow: Supervisor → Search (DDG + arXiv + Wikipedia) → Code (Python in sandboxed Docker) → Analysis → Skeptic → back to Supervisor if the conclusions don't hold up.

The focus is not just on execution, but on validation — the system actively tries to break its own conclusions via a skeptic loop.

Stack: LangGraph, LangChain, FastAPI, ChromaDB, SQLite (before that, there was PostgreSQL and Redis), nginx. Works with Ollama (local/cloud) and any OpenAI-compatible API. One-command Docker Compose setup.

https://github.com/Evidion-AI/EvidionAI

I'm open-sourcing this because a lot of similar projects are emerging right now, each exploring different approaches to autonomous research and agent orchestration.

I'm building this solo, and it seems more valuable to share the approach early rather than develop it in isolation — especially since the space is evolving fast and there’s clearly convergent interest in this direction.

Would really appreciate feedback on the agent architecture — especially the Supervisor routing logic, which is currently the most fragile and interesting part.