Worked on the AMD-AGI/Primus repository to deliver two targeted improvements for high-performance training workflows. Addressed network reliability by implementing RDMA adapter filtering, ensuring that only appropriate InfiniBand HCAs and network interfaces are selected while excluding GPU and non-socket devices. Enhanced job observability by updating Kubernetes pretraining scripts to redirect both stdout and stderr to log files, guaranteeing log directory creation and more robust output tracking. These changes, developed using Shell scripting and leveraging expertise in Kubernetes and networking, reduced the risk of misconfiguration and missing logs, contributing to more stable and debuggable distributed system operations within the project.