The production LLM router achieved eight‑fold more even memory distribution across its backend instances compared to the author's implementation. However, this approach resulted in failure rates five times higher than the baseline.