Worked on the NVIDIA-NeMo/Megatron-Bridge repository to deliver enhancements for Gemma 3 Rotary Positional Encoding, focusing on aligning rotary embeddings with a new tensor structure. Leveraged deep learning and machine learning expertise, primarily using Python and PyTorch, to improve tensor handling and ensure compatibility for both training and inference scenarios. The implementation reduced setup complexity and increased model stability, enabling smoother experimentation with RoPE-based models. Emphasized code hygiene through signed-off commits and clear pull request references, supporting traceability and maintainability. This work laid the foundation for future performance optimizations within the Megatron-Bridge integration, with attention to robust unit testing practices.