Worked on improving ONNX export reliability in the pytorch/pytorch repository, focusing on transformer model compatibility. Addressed a critical issue in the ONNX Attention-23 exporter by correcting how the number of attention heads is inferred, ensuring the exporter reads the correct input dimension. This fix restored stable ONNX export for Attention-23, reducing shape-related export failures and improving interoperability with ONNX runtimes. Demonstrated strong debugging and dimension analysis skills within the PyTorch codebase, using Python and deep learning frameworks to enhance code maintainability. The work contributed to more robust model export pipelines, supporting machine learning workflows that rely on accurate ONNX conversion.