Worked on optimizing the RT-DETR model architecture within the ultralytics/ultralytics repository by refactoring the RepC3 module to adjust the output channels of convolutional layers, which improved overall model performance. Addressed a critical issue in the RepC3 module that affected stability and reliability, ensuring more robust behavior for RT-DETR models. The refactor also enhanced the modularity of the codebase, supporting future enhancements and easier iteration. Leveraged deep learning and computer vision expertise, primarily using PyTorch and Python, to deliver these improvements. The work focused on both architectural optimization and code maintainability within a production-grade deep learning framework.