Show HN: Gittins, an online bandit engine that enforces log correctness (www.getgittins.dev)

Gittins, a new online contextual bandit engine, has been introduced as a practical solution for real-world decision-making problems. Designed to optimize performance in production environments, Gittins allows for simultaneous learning from observations in constant time and memory while expecting non-stationarity in data—critical for applications where user behavior evolves over time. It employs a simple epsilon-greedy algorithm that maintains exploration throughout its operation, addressing challenges inherent in dynamic environments such as varying action sets and delayed reward feedback. The significance of Gittins lies in its robust design that encourages safe reward handling and provides a transparent means of logging decision-making processes to ensure correctness. Its architecture, which allows features to be learned without prior schema registration and supports flexible deployment options (in-memory, file-based, or via shared services), makes it adaptable for a wide range of applications. The engine’s ability to swap prediction models and actions dynamically, coupled with rigorous testing for performance consistency across platforms, positions Gittins as a valuable tool for developers in the AI/ML community looking for a production-ready solution in contextual bandit problems.