Worked on the nvidia-holoscan/holohub repository to enhance the stability and reliability of the stereo vision pipeline, focusing on memory management during rectification parameter setup. Addressed a critical segmentation fault by refactoring the handling of CUDA memory, specifically removing an unnecessary cudaMemcpy and correcting the cudaFree operation for the d_P pointer. This fix improved real-time performance and reduced production incidents by ensuring proper memory lifecycle management. Utilized C++ and CUDA to deliver a targeted solution that lowered runtime crashes and support overhead, contributing to more maintainable stereo processing code and a smoother deployment experience for the stereo vision system.