Tiny Router

Pre-classify every message in <10ms with a local DeBERTa-v3-small ONNX model (44M params). Routes low-value messages (thanks, ok, got it) away from the LLM — saving 30-50% on API costs at zero latency cost.