Worked on the h2oai/h2o-3 repository to enhance the reliability of XGBoost MOJO pipelines by addressing a critical issue in prediction logic. Focused on Java development and model deployment, the work involved fixing a bug to ensure zero offsets in predictions are handled correctly, improving the accuracy of models trained with or without offsets. Comprehensive tests were added in Java, along with new MOJO test data, to validate behavior across different training scenarios. This approach strengthened test coverage and reduced the risk of offset-related errors in production, demonstrating attention to correctness and robustness in machine learning model integration and testing.