During June 2025, Chaojhou Hou enhanced the AMD-AGI/Primus repository by delivering two targeted features focused on network reliability and job observability for high-performance training workloads. Chaojhou implemented RDMA adapter filtering, ensuring that GPU and non-socket interfaces were excluded from InfiniBand HCA selection, which reduced the risk of misconfiguration in RDMA networking. Additionally, Chaojhou improved the Kubernetes pretraining workflow by redirecting job output streams to structured log files and automating log directory creation, strengthening debugging and monitoring capabilities. The work demonstrated depth in Kubernetes orchestration, advanced networking, and shell scripting, resulting in more robust and maintainable system administration practices.