Worked on performance and reliability improvements for Kokoro workloads in the openvinotoolkit/openvino and aobolensk/openvino repositories, focusing on NPU-based inference optimization and pipeline refinement. Leveraged C++ and deep learning expertise to implement NPUW NONE partitioning, optimize LSTM input handling, and streamline model caching, resulting in reduced latency and increased throughput for real-time AI workloads. Collaborated across repositories to align parameter handling and caching strategies, and validated changes with benchmark-driven testing. Additionally, contributed to codebase maintainability by removing obsolete DownsampleInterpolate workarounds, simplifying the interpolation-downsampling logic, and reducing maintenance overhead, supporting safer future enhancements and improved onboarding for new contributors.