Worked on the CogitoNTNU/DeepTactics-Muzero repository to improve code observability and correctness within the training loop. Focused on backend development and debugging, the work involved enhancing training loss logging by tracking total policy, reward, and value losses, with an option to display aggregated losses during training for easier debugging. Addressed a duplicate calculation in the calculate_loss function and clarified the initialization and accumulation of total losses to prevent potential logging errors. Used Python and applied deep learning and machine learning principles to strengthen maintainability, enabling faster debugging and more reliable interpretation of training metrics throughout the development process.