Worked on the Verl-deepresearch repository to enhance reliability and observability, focusing on backend stability rather than new feature development. Addressed a critical issue in MLflow metric logging where metric names containing special characters, specifically the '@' symbol, caused failures in experiment tracking. Implemented a solution in Python that sanitized metric names by replacing '@' with '_at_', ensuring compliance with MLflow’s naming constraints and enabling consistent data tracking across experiments. This fix improved maintainability and third-party integration, allowing for reliable experiment metrics collection. The work demonstrated attention to detail in data tracking, MLflow integration, and Python-based backend development.