Ecotopia

Political eco-simulation game. You are the mayor of a city facing ecological collapse: 7 rounds, tile-based grid, three resource bars (ecology, economy, research). Deliver free-text speeches, watch AI extract your promises, and face citizen reactions powered by fine-tuned Mistral models.

Team: Nolan Cacheux, Kyle Kreuter Built at: Mistral Hackathon London 2026 (48 hours)

What It Does

Ecotopia puts you in charge of a dying city through 7 rounds of ecological crisis. Each round, you enact policies on a tile-based grid (close a coal plant, fund research, build infrastructure) and deliver a free-text speech to your citizens. Fine-tuned models extract promises from your words and track them across rounds. Citizens react based on their personality, profession, and memory of your past speeches. Broken promises erode trust; honesty builds it.

Architecture

Speech → FT-Extract (12B) → Contradiction Detection → FT-Citizens (8B) → ElevenLabs TTS → UI

Stage	Description
Speech Input	Player types a free-text speech
FT-Extract	Fine-tuned Mistral Nemo 12B extracts promises with deadlines and detects contradictions (JSON)
Game Engine	Deterministic state updates: ecology, economy, research scores
FT-Citizens	Fine-tuned Ministral 8B generates citizen reactions, dialogues, and approval deltas
TTS	ElevenLabs gives each citizen a unique voice (10 voice profiles)
UI	Phaser 3 pixel art frontend with tile grid, citizen sprites, and speech bubbles

Tech Stack

Layer	Technology
Frontend	Phaser 3, TypeScript, Vite, Pixel Art
Backend	Spring Boot 3.5, Spring AI, PostgreSQL
AI Models	Fine-tuned Mistral via QLoRA on HuggingFace Inference Endpoints
TTS	ElevenLabs Turbo v2.5 (unique voice per citizen)
Monitoring	Weights & Biases (training metrics + evaluation)

Fine-Tuned Models

Model	Task	Base	HuggingFace
ecotopia-extract-ministral-8b	Promise extraction	Ministral 8B Instruct	LoRA / Merged
ecotopia-extract-nemo-12b	Promise extraction	Mistral Nemo 12B	LoRA
ecotopia-citizens-ministral-8b	Citizen reactions	Ministral 8B Instruct	LoRA / Merged
ecotopia-citizens-small-22b	Citizen reactions	Mistral Small 22B	LoRA / Merged

Datasets:

• ecotopia-extraction-dataset (300 examples)
• ecotopia-citizens-dataset (390 examples)

Setup

Prerequisites

• Java 21+
• Docker (for PostgreSQL)
• Node.js 18+ (for frontend)

Quick Start

# 1. Copy environment variables
cp .env.example .env
# Edit .env with your API keys

# 2. Start database
cd backend && docker compose up -d

# 3. Start backend (port 7777)
cd backend && ./mvnw spring-boot:run

# 4. Start frontend (port 5173)
cd frontend && npm install && npm run dev

Open http://localhost:5173 in your browser.

Environment Variables

Variable	Required	Description
EXTRACTION_ENDPOINT	Yes	HuggingFace endpoint URL for 8B extraction model
CITIZEN_ENDPOINT	Yes	HuggingFace endpoint URL for 24B citizen model
HF_TOKEN	Yes	HuggingFace API token (for endpoint auth)
ELEVENLABS_API_KEY	No	ElevenLabs API key (enables citizen voice TTS)
DB_USERNAME	No	PostgreSQL username (default: ecotopia)
DB_PASSWORD	No	PostgreSQL password (default: ecotopia)

API Endpoints

Method	Path	Description
POST	/api/games	Create a new game
GET	/api/games/{id}	Get game state (resources, citizens, tiles)
POST	/api/games/{id}/speech	Submit speech (triggers AI extraction + reactions + TTS)
POST	/api/games/{id}/end-round	End current round, advance to next
GET	/api/games/{id}/promises	List all tracked promises
POST	/api/games/{id}/tiles/{tileId}/action	Execute tile action

Speech Response Format

{
  "extractedPromises": [
    {"id": 1, "text": "build wind farms", "deadline": 3, "status": "ACTIVE"}
  ],
  "contradictions": [
    {"description": "Promises forests while building coal plants", "severity": "high"}
  ],
  "citizenReactions": [
    {
      "citizenName": "Karl",
      "dialogue": "Wind farms won't replace factory jobs.",
      "tone": "suspicious",
      "approvalDelta": -3,
      "audioBase64": "base64_encoded_mp3..."
    }
  ]
}

HuggingFace Endpoints

Both fine-tuned models are deployed on HuggingFace Inference Endpoints with NF4 4-bit quantization via custom handler.py. Endpoints scale to zero after 15 minutes of inactivity.

Cold start warning: First request after idle takes ~5 minutes. The backend retries automatically (10x with 30s delay).

W&B Report

Full training metrics, evaluation benchmarks, and model comparisons:

Ecotopia: Fine-Tuning Mistral for Political Simulation (W&B)

PDF Report

License

MIT