: Prototype with XGBoost + ONNX Runtime. If model size exceeds limits, prune trees or fall back to LightLTR. Avoid cloud APIs unless you cannot run inference on-device at all.