The Proposed Architecture My team and I propose separating the encoder from the rest of the model architecture: 1. Deploy a lightweight encoder on the wearable device's APU (AI processing unit). 2.