Justin Rosner contributed to ROCm/rocMLIR and ROCm/llvm-project by developing and refining MLIR transformations for convolution operations, including enhanced support for depthwise and backward data convolutions. He refactored core components to improve maintainability and extensibility, and addressed correctness in AMDGPU folding logic to prevent faulty optimizations. Justin expanded end-to-end testing with hardware-aware gating, ensuring robust CI coverage for MLIR-based pipelines. He also introduced new tensor load and store operations in the ROCDL dialect, enabling efficient data movement between global memory and Local Data Share. His work leveraged C++, MLIR, and LLVM IR, demonstrating depth in compiler development and GPU programming.