During October 2025, Diotima developed a fused PolyNorm operator for the linkedin/Liger-Kernel repository, focusing on deep learning performance and memory optimization. Leveraging CUDA kernels and the Triton language, Diotima implemented both forward and backward passes within a single operator, enabling efficient training workflows. The approach included a naive PyTorch implementation for validation, ensuring correctness and providing a baseline for comparison. This work achieved a 12-40x speedup and reduced memory usage by approximately 6.4 times compared to standard PyTorch autograd. The depth of engineering demonstrated strong proficiency in CUDA, PyTorch, and Triton, addressing key bottlenecks in model training.