AI Poker Model Evaluation Game

A competitive Texas Hold'em poker game where 5 AI models compete against each other. Built for Vercel's "Build your own model eval game" hackathon.

Features

• Two Game Modes:

  • Fast Mode: Uses smaller, faster models (GPT-4o-mini, Claude Haiku, Gemini Flash, Grok-2, etc.)
  • Smart Mode: Uses larger, more capable models (GPT-4o, Claude Sonnet, Gemini Pro, Grok-2, etc.)

• Real-time Gameplay: Watch AI models make decisions in real-time

• Complete Game Logging: Every action is logged with full context

• Replay System: Review games at variable speeds, step through actions

• AI Reasoning: Click any action to see the model's reasoning (fetched asynchronously)

Setup

1. Install dependencies:

bun install

2. Set up environment variables in .env.local:

OPENAI_API_KEY=your_key_here
ANTHROPIC_API_KEY=your_key_here
GOOGLE_GENERATIVE_AI_API_KEY=your_key_here
GROK_API_KEY=your_key_here
# Or use XAI_API_KEY as alternative

3. Run the development server:

bun dev

4. Open http://localhost:3000 in your browser.

Deployment

Deploy to Vercel:

vercel

Make sure to add all API keys in Vercel's environment variables settings.

Architecture

• Poker Engine (lib/poker/): Complete Texas Hold'em game logic
• AI Integration (lib/ai/): Vercel AI SDK integration with 5 providers
• API Routes (app/api/game/): Game state management and AI action processing
• UI Components (components/): React components for poker table visualization

Performance Optimizations

• Action-only prompts (reasoning fetched async)
• 15-second timeout per decision
• Streaming UI for "thinking" indicators
• Pre-recorded demo games for instant first-visit experience

Game Flow

1. User selects Fast or Smart mode
2. 5 AI players start with 1000 chips each
3. Game continues until one player has all chips
4. All actions logged for replay
5. Final leaderboard with rankings