Mirofish Offline Simulation

This is a fully local multi-agent simulation engine that generates hundreds of AI personas to predict how people will react to documents like press releases or policy drafts. It uses Neo4j for graph memory and Ollama for the LLM layer, so nothing hits cloud APIs. You feed it a document, it extracts entities and relationships, spawns agents with distinct personalities, then simulates them posting, arguing, and shifting opinions over simulated hours on platforms like Twitter or Reddit. The Python API is clean and the hybrid search combines vector similarity with BM25. Requires serious hardware though: the recommended qwen2.5:32b model needs 24GB VRAM, so you'll want to drop down to the 14b variant unless you have a beefy GPU.