TinyIntent
Voice commands spoken on an iPhone, carried out on a Mac. Shortcuts handles the dictation. A small ELECTRA classifier, exported to Core ML and accelerated on the Apple Neural Engine, decides whether a request is handled locally or escalated. Helpers cover weather, system metrics, network telemetry, log analysis, and trading. The Mac bridge sits behind a shared secret, with a capped request size and a whitelisted label set. I trained the classifier myself on 439 labeled examples.
Highlights
01On device classifier. No cloud round trip for routing.
02ELECTRA small backbone, 13.5 million parameters, exported to Core ML for the Neural Engine.
03Helpers are sandboxed. Schema validated, with two step confirmation for critical actions.
04Hardened bridge. Shared secret auth, request size caps, label whitelist, LAN exposure opt in.
05Multi domain. Weather, system, network, trading, and DevOps from one voice interface.
Repository
evanacci/tinyintentpublic
Stack
Swift,CoreML,Python,FastAPI,PyTorch,Ollama