LLM Bidding Strategy Arena 🤖💸

An AI-powered system that generates, validates, and evaluates Real-Time Bidding (RTB) strategies using Large Language Models (Gemma3 via Ollama).

🌟 Features

• Automated Strategy Generation: Uses LLMs to write Python bidding code based on high-level goals (e.g., "Maximize Impressions", "Aggressive").
• Sandboxed Execution: Validates generated code for safety and syntax.
• Historical Replay Engine: Simulates auctions using mock or real historical data.
• Interactive Arena: Compare multiple strategies side-by-side with real-time visualization.
• Design Patterns: Built with Clean Architecture principles (Strategy, Builder, Adapter, Factory).

🚀 Quick Start

1. Prerequisites

• Python 3.8+
• [Optional] Ollama installed and running with gemma3:12b model for real AI generation.
  • Run ollama run gemma3:12b to pull the model.

2. Installation

pip install -r requirements.txt

3. Run the Arena

You can launch the web interface using the provided helper script:

chmod +x run_arena.sh
./run_arena.sh

Or manually:

streamlit run bidding_arena/visualization/app.py

🏗 System Architecture

• core/: Interfaces and base logic (Clean Architecture).
• generation/: LLM interaction, Prompt Builder, and Code Validator.
• data/: Data generation and loading.
• visualization/: Streamlit dashboard.

🛠 Design Patterns Used

• Strategy Pattern: IBiddingStrategy allows swapping different bidding logic (generated code) at runtime.
• Builder Pattern: PromptBuilder constructs complex prompts based on strategy configuration.
• Adapter Pattern: ILLMClient abstracts the AI provider (Mock vs Ollama).
• Factory Method: StrategyGenerator encapsulates the creation and validation of strategy objects.

🧪 Testing

Run unit tests (coming soon):

pytest tests/