I Built an AI Simulation Engine You Talk to Like ChatGPT — Meet MiroFish

#ai #webdev #typescript #showdev

What if you could rehearse a crisis before it happens, pressure-test a campaign before you spend, or map the second-order reactions to a pricing change — all by asking a question in plain language?

That's MiroFish.

The Problem: Why ChatGPT Isn't Enough

Ask ChatGPT for a prediction, and it collapses competing audience reactions into one confident answer. But real-world outcomes depend on people reacting to people — shifting incentives, emergent narratives, and resistance clusters that a single summary paragraph can't capture.

So I built something different.

Architecture: Multi-Agent Simulation in a Chat Interface

MiroFish runs a 5-stage pipeline behind a single chat conversation:

User Question → Knowledge Graph → Agent Simulation → Report → Follow-up Queries

Stage	What Happens
1. Seed	Plain-language question + optional file uploads (PDF, Markdown) for grounding
2. Graph	LLM extracts actors, relationships, pressures, and factual anchors into a structured knowledge graph
3. Simulate	AI personas interact across short-form and threaded social surfaces over multiple rounds — each with persistent memory and distinct incentives
4. Report	Structured result card with executive summary, risk signals, narrative paths, and confidence indicators
5. Query	Keep asking the simulated world — unlike a static forecast, this is interactive

The key insight: agents don't just answer. They react to each other, creating emergent behavior that a flat prompt-response loop can't produce.

What You Can Use It For

Use Case	Example
Campaign Testing	"What happens if we launch in a skeptical category?"
Pricing Reactions	"If we raise prices, which segments push back first?"
Policy Stress Tests	"Where does support split when this goes public?"
Market Narratives	"What if positive news meets coordinated skepticism?"