On-device agent: a small LLM picks an ExecuTorch vision model, reads the result, and answers - View it on GitHub