Worked on the JuliaGPU/CUDA.jl repository to enhance the reliability of sparse matrix operations, focusing on the integrity of Block Sparse Row (BSR) descriptors. Addressed a critical bug by correcting how the number of non-zero elements and matrix dimensions were passed to the cusparseCreateBsr function, ensuring accurate descriptor creation for BSR matrices. This fix, implemented in Julia using CUDA and GPU computing techniques, improved the stability and correctness of sparse linear algebra routines. The work reduced downstream errors and debugging time for users relying on CUDA.jl’s sparse matrix support, reflecting a deep understanding of GPU-based numerical computing challenges.