Worked on the vllm-project/tpu-inference and vllm-project/ci-infra repositories to enhance resource management and CI/CD workflows for Google Cloud TPU environments. Improved the Gemma4 model’s resource handling by correcting data-parallel-size configuration and introducing explicit shutdown methods, which reduced test latency and improved reliability. Developed persistent JAX compilation cache storage in Google Cloud Storage to accelerate TPU inference testing. In the CI/CD pipeline, applied Terraform-based infrastructure updates, streamlined JAX cache setup, and enabled secure access to private repositories using deploy keys and SSH. Utilized Python, Terraform, and Bash to deliver maintainable, secure, and efficient solutions for machine learning infrastructure.