Developed and delivered a performance-focused feature for the InfiniCore repository, implementing efficient inference with a paged key-value cache for multi-head attention mechanisms. This work introduced a paged KV cache operation in C++ and CUDA, enabling single-step decoding while reducing memory usage and improving inference throughput for neural network attention layers. The feature addressed memory efficiency challenges in machine learning inference by optimizing cache management for attention models. The development process included end-to-end ownership, from design through code-level delivery, with clear traceability to tracked issues. The work demonstrated expertise in C++ development, CUDA programming, and neural network architecture optimization.