A college student built a SimCity for AI forecasting — 18k stars in 4 months

You know that feeling when every prediction model you've tried treats human behavior like a physics equation — inputs in, outputs out — and completely misses the part where one tweet triggers a cascade, where a minority opinion suddenly becomes a majority, or where two factions merge and flip the outcome? Before MiroFish, your options for social-dynamics forecasting were: run statistical regression (ignores interactions), use an ML model trained on historical data (can't generalize to novel events), or hire an analyst to write a qualitative scenario report (expensive, slow, not reproducible). Now: you upload seed material, describe your prediction question in natural language, and an engine spawns hundreds of agents who argue it out — then hands you a structured report from the emergent behavior.

You upload 'seed material' — a news article, policy document, financial report, or even fiction — and MiroFish first builds a knowledge graph from it using GraphRAG, extracting the key players, relationships, and tensions. From that graph, it automatically generates agent personas: each gets a unique personality, a stance on the topic, and long-term memory powered by Zep Cloud. Those agents then run on OASIS (an open-source simulation engine by CAMEL-AI that scales to 1 million agents) across two parallel social environments — one Twitter-like, one Reddit-like — where they post, argue, persuade, and update their memories as events unfold. When the simulation ends, a dedicated ReportAgent synthesizes the emergent behavior into a structured prediction report. The clever design insight: predictions emerge from agent behavior, not from the model's training data — so MiroFish can forecast novel events that never happened before, as long as the human dynamics are plausible.

• GraphRAG knowledge extraction — instead of treating your input document as flat text, MiroFish builds a structured map of who the players are and how they connect, giving agents a grounded understanding of the situation...
• Auto-generated agent personas with long-term memory — each simulated agent gets a unique personality, perspective, and Zep Cloud-backed memory that persists across the simulation, so agents remember earlier arguments an...
• OASIS simulation engine underneath — the multi-agent interaction layer scales to 1 million agents and supports 23 social actions (follow, reply, repost, like, etc.), meaning your simulation resembles the actual dynamics...
• Dual-platform parallel simulation — agents run simultaneously on both a Twitter-like and a Reddit-like environment, capturing both short-form viral dynamics and long-form discussion patterns in the same prediction
• Natural language prediction questions — you describe what you want to forecast in plain English ('will this policy face significant public backlash within 30 days?') and the system tracks that question throughout the si...
• God's-eye variable injection — you can pause the simulation and inject new events mid-run ('what if a major news outlet publishes a counter-narrative on day 3?'), letting you run multiple scenario branches without resta...
• One-command full-stack setup — npm run setup:all installs all Python and Node dependencies, creates the virtual environment, and launches frontend at localhost:3000 and API at localhost:5001, reducing setup to under 5 m...

Who It Is For

If you work on public opinion analysis, policy forecasting, social media intelligence, or scenario planning — and you're frustrated that every existing tool either drowns in data or ignores emergent group dynamics — MiroFish gives you a framework to simulate the messy human part. It's also genuinely interesting for AI researchers studying multi-agent emergent behavior. Not production-ready for en...

Worth Exploring?

Yes — the architecture here is genuinely clever and the 18k star count in under four months from a solo student developer is a real signal. The use case for simulating public opinion around policy or PR scenarios is immediately obvious and commercially underserved. The honest caveat: you're running at v0.1.0 with 32 open issues and the simulation quality is directly proportional to how well you craft your seed material and agent setup — garbage in, garbage out applies doubly here because bad seed data produces confident-sounding but wrong agent personas.