A modular, config-driven framework to: - ingest PyTorch or ONNX models, - convert them into Qualcomm QNN/QAIRT artifacts and context binaries, - profile runtime on device targets (mock backend included, QAI Hub backend scaffolded), - iterate optimization decisions with a LangGraph control loop until latency targets are met.