Worked on the NVIDIA-NeMo/Gym repository to improve the reliability and data integrity of the metrics pipeline, focusing on backend systems under high concurrency. Addressed a critical bug by introducing a centralized file locking utility and implementing atomic write operations in Python, replacing the previous directory-based locking approach. This solution prevented JSON decoding errors and data corruption during simultaneous metrics updates, particularly for SWE agent metrics. The work emphasized robust file I/O and concurrency management, enhancing system resilience and preparing the pipeline for larger-scale, concurrent workloads. Delivered a single, well-documented commit that improved job stability and safeguarded metrics storage processes.