Interesting approach! How does this handle rate limiting and token costs at scale? Would love to see benchmarks.