Most LLM routers are Python wrappers or cloud proxies that add tens of milliseconds before a request reaches the model. Jev-router is an in-process, zero-allocation native Rust engine that classifies and assigns routes in under 5 ms. Its deterministic fallback cascade uses byte-level automata, subword classifiers, and closed-form evaluators before escalating to an embedding model or LLM. Simple requests are resolved locally, without touching either.