Why I chose on-device inference for a writing assistant

When I built DraftKey, the biggest decision was running the model locally instead of calling a cloud...