Worked on the openvinotoolkit/openvino repository to implement LFM2 model support and stateless processing, targeting improved deployment on NPU architectures. The engineering effort focused on adapting model naming conventions, updating convolution processing logic, and modifying ReadValue and Assign pathways to enable stateless input handling. This transition from stateful to stateless models reduced runtime latency and memory usage, aligning with the LFM2 architecture’s requirements of ten short LIV convolution blocks and six grouped query attention blocks. The work leveraged C++ programming, NPU development, and model optimization skills, contributing a core feature that enhances efficiency and compatibility for machine learning workloads.