Worked on the OpenCV repository to address a critical bug in the VBLAS SIMD implementation, focusing on improving numerical stability and preventing out-of-bounds storage in high-performance computation paths. Applied SIMD programming and numerical methods in C++ to introduce fused multiply-add (FMA) operations, which enhanced precision and reduced the risk of overflow or underflow. Simplified the reduction path to optimize throughput and lower computation overhead, resulting in more efficient and robust math routines for computer vision workloads. The solution was backported to the OpenCV 4.x branch, ensuring broader reliability and consistency across platforms for high-performance numerical applications.