Worked on the vllm-project/tpu-inference repository to deliver a profiling performance and resource management optimization for TPU inference workloads. The solution limited TPU tracing to a single core and tile, reducing profiling overhead and stabilizing resource usage. By narrowing the tracing scope, the approach enabled faster profiling cycles and clearer performance signals, supporting more targeted optimizations and smoother development iteration. The work involved Python development with a focus on performance tuning, profiling, and TPU optimization. Changes were managed through commit-based workflows and cross-team collaboration, preparing the feature for release to enhance performance visibility and support future profiling improvements.