That sounds like a neat little experiment—old phone, neural net, huh? If you’ve got a lightweight model and a good quantization setup, I can probably help you squeeze a bit more performance out of the hardware. Let me know what framework you’re using and what you’re aiming to detect or classify. Maybe we can shave off some cycles with a custom kernel or tweak the batch size. Just ping me the specs and the code, and I’ll dig into it.