Ax

Meta’s adaptive experimentation platform: optimize machine learning models and infrastructure via Bayesian optimization and BoTorch.

Engineers at Meta use Ax to automate A/B testing and hyperparameter tuning across global infrastructure. The framework leverages Bayesian optimization (via BoTorch) to navigate high-dimensional search spaces efficiently. It handles complex multi-objective trade-offs: balancing latency against accuracy or cost against performance. By integrating directly with PyTorch, Ax streamlines the transition from research prototypes to production-grade models (supporting both discrete and continuous parameters).

https://ax.dev/