Worked on the rapidsai/rmm repository to enhance documentation around device_buffer synchronization semantics, focusing on multi-stream CUDA workloads. The update clarified that set_capacity, resize, and reset operations do not synchronize CUDA streams, emphasizing the need for caller-managed synchronization before allocation to ensure data validity. This documentation improvement, implemented using C++ and technical writing skills, aimed to reduce the risk of data races and misuses without altering the existing API or introducing code changes. By providing explicit guidance on synchronization requirements, the work improved safety and onboarding for developers working with device_buffer in complex, concurrent GPU programming environments.