During May 2025, contributed a performance-focused feature to the NVIDIA/NeMo-RL repository by reworking worker initialization to operate asynchronously and in parallel. Leveraging Python, Ray, and asynchronous programming techniques, the new approach introduced a create_worker_async method that returns futures, allowing worker creation tasks to proceed concurrently. The initialization flow was updated so that results are awaited in batches, with RayWorkerGroup collecting and resolving all worker references simultaneously using ray.get. This redesign reduced startup latency and improved scalability for distributed reinforcement learning workloads, optimizing resource utilization and enabling faster cluster readiness for large-scale training scenarios without introducing new bugs.