Worked on the ml-explore/mlx repository to address a critical issue in the Post-LN TransformerDecoderLayer, focusing on the correct routing of cross-attention queries. Used Python and deep learning techniques to ensure the appropriate tensor was utilized during attention calculations, directly improving model reliability. Developed and integrated comprehensive tests for various Transformer configurations, which enhanced robustness and facilitated early detection of edge cases. Emphasized maintainability by providing clear commit messages and expanding test coverage, supporting future development. The work demonstrated a methodical approach to debugging and stabilization within neural network architectures, contributing to the overall reliability of machine learning workflows.