Worked on integrating the REMAX advantage estimator into the microsoft/agent-lightning repository by refactoring the trainer to use an asynchronous rollout manager for sequence generation. This approach replaced direct calls to actor_rollout_wg.generate_sequences with self.async_rollout_manager.generate_sequences, resulting in a more scalable and resilient sequencing pipeline. The refactor improved throughput and fault isolation, laying the foundation for advanced estimator capabilities. Demonstrated strong skills in Python, reinforcement learning, and software engineering, with a focus on asynchronous programming and modular pipeline design. The work addressed the need for robust sequence generation, supporting future enhancements and enabling more efficient reinforcement learning workflows.