Building a smart routing agent that picks the cheapest model for every task — minimizing tokens without sacrificing accuracy on AMD GPUs.