// ls /ideas

Peer-to-Peer Research Orchestrator

A decentralized research network where each participating device acts as an LLM inference node. Nodes run local models via Ollama and are coordinated through a central GitHub repository that serves as both a task registry and a knowledge base.

• nodes — each device runs Ollama locally; no cloud GPU required. Heterogeneous hardware is fine — bigger models on beefier machines, smaller models on edge devices.
• github_as_db — the repo holds a task_queue.json (open research topics) and a node_registry.json (which node is responsible for what). Nodes poll / webhook on changes and self-assign tasks.
• knowledge_sync — nodes periodically push findings (summaries, embeddings, citations) back to the repo via PRs or direct commits to a knowledge/ directory. Merge = consensus.
• long_running_tasks — heavy tasks (literature review, multi-step reasoning chains) are broken into subtasks in the queue and delegated across available nodes. GitHub Actions can act as an orchestrator layer for sequencing.
• coordination — no central server beyond GitHub. Conflict resolution happens through git history; audit trail is free.

// status: draft — ideas welcome