Worked on the PrimeIntellect-ai/prime-rl repository to enhance the reliability of checkpoint management during long-running training runs. Addressed a recurring NotADirectoryError by implementing a safer deletion process for older checkpoints, replacing file-based removal with directory-wide cleanup using Python’s shutil.rmtree. Focused on backend development and robust file system operations, the solution reduced manual intervention and improved automation safety. The targeted cleanup logic not only prevented errors but also supported future scalability and maintainability of the codebase. Emphasized thorough testing and careful handling of filesystem edge cases, ensuring that checkpoint deletion is both reliable and adaptable for evolving project requirements.