Self-hosted agentic research pipeline

Overview

An autonomous multi-agent research pipeline that runs open-weight language models on a campus GPU workstation through Ollama and is driven from a laptop over Tailscale. Nine playbooks cover literature review, ideation, model refinement, proof attempts, simulation experiments, draft updates and adversarial verification.

How it works

Eight specialized agents are orchestrated with LangGraph state machines using mixture-of-agents ensembling: every step queries three local models (Gemma 26B, DeepSeek-R1 32B, Qwen 27B) and merges their outputs with a cloud synthesizer model. A token-authenticated FastAPI server and browser dashboard provide live log streaming over SSE, pause/resume, human-in-the-loop approval checkpoints and SQLite checkpointing for crash-resumable runs. All artifacts are committed append-only to a git-backed research knowledge base.

Tools: Python, LangGraph, FastAPI, Ollama, SQLite, Tailscale.