Developed and integrated a Triton-based depthwise 2D convolution kernel for the Iluvatar backend within the FlagOpen/FlagGems repository, focusing on optimizing depthwise operations for specific kernel sizes. The work bypassed generic convolution paths to deliver measurable performance improvements in computer vision inference workloads. Leveraging expertise in backend development, kernel optimization, and technologies such as Python, PyTorch, and Triton, the developer ensured the new operator was registered for immediate production use. The contribution included rigorous code practices and seamless backend integration, providing a foundation for broader performance gains and demonstrating a strong focus on practical, business-driven engineering outcomes.