Worked on the keras-team/keras-io repository to enhance the modularity and performance of deep learning tutorials and benchmarks. Refactored the model definition into a function-based approach using Python and Keras, improving code reusability and maintainability. Integrated gradient centralization into the training loop to boost training stability and performance, addressing issues in existing examples to ensure reliability. Updated documentation and Jupyter notebook execution counts to accurately reflect these architectural changes, supporting reproducibility for users. The work focused on deep learning and machine learning workflows, delivering a single feature that improved both the structure and efficiency of the repository’s training processes.