Worked on performance optimization for TPU inference pipelines in the vllm-project/tpu-inference repository, focusing on cache selection during inference. Developed and integrated a JIT-compiled function using JAX and Python to select key-value caches based on indices, which improved throughput and reduced latency for TPU-based machine learning workloads. The solution was closely aligned with the Qwen 3.5 integration and included clear traceability through signed-off commits and collaborative work with other contributors. This work demonstrated depth in performance tuning, caching strategies, and cross-team collaboration, contributing to more efficient and maintainable TPU inference processes within the machine learning infrastructure.