This project is a full-stack, end-to-end LiveOps dashboard designed to automate and optimize the decision-making process for in-game content releases. It is intended to be a data sandbox and a demonstration of how machine learning (ML) can improve upon manual A/B testing, driving superior business outcomes like higher tutorial completion rates.
The tool allows a user to set up content variants, create a new campaign, and then run simulations of various machine learning algorithms to optimize tutorial completion rate.
The platform dynamically allocates traffic to the best-performing content variants as the simulation progresses, focusing on the core Multi Armed Bandit principle of exploration vs. exploitation.
Check out a video here: https://www.youtube.com/watch?v=vEyCZdIfBWY
Frontend: React, Vite, Tailwind CSS, Recharts
Backend: Python, FastAPI, SQLite
ML Algorithms: Multi-Armed Bandit (MAB), Segmented MAB. Algorithms implemented from scratch in Python (no machine learning libraries).
Key Features and Capabilities
-
Dynamic Campaign Management: Users can configure and launch new simulated campaigns, defining content variants, setting simulation parameters, and specifying ground-truth performance.
-
Automated Optimization Strategies: The tool implements and visualizes the performance of two distinct machine learning strategies:
- Standard Multi-Armed Bandit (MAB): Dynamically identifies and exploits the single highest-performing content variant.
- Segmented MAB: Applies unique MAB policies to distinct, user-defined segments (e.g., region, platform), each with modifiers to tutorial completion rate, maximizing performance across a heterogeneous user base.
-
Data-Driven Visualization: A modern frontend provides a clear, real-time dashboard view of campaign performance, showing variant traffic allocation, cumulative rewards, and overall metrics.
-
API-First Design: The simulation engine is built with an API-first approach, allowing for easy extension into a production environment where real-world products could hook into the platform's decision-making service.
