Worked on advanced model training and reinforcement learning workflows across two repositories. In nvidia-cosmos/cosmos-transfer1, delivered distillation training support by implementing configuration, model architecture, dataset handling, and inference utilities, enabling efficient single-step distillation of large models using PyTorch and Hydra. For NVIDIA-NeMo/Gym, integrated the Harbor agent to support multi-turn reinforcement learning workflows, developing a full integration package with custom agent and environment wiring, rollout conversion, and comprehensive documentation. Enhanced onboarding and reproducibility through improved tests and CI coverage. The work demonstrated depth in distributed training, Docker, and full stack development, focusing on reliability and extensibility for future features.