evanacci.dev / project / tinyintent
↳ project 03 in production

TinyIntent

“voice automation, local first”

TinyIntent is a voice automation platform that respects your privacy. iPhone Shortcuts dictates your speech, then a tiny DistilBERT classifier, exported to Core ML and accelerated on the Apple Neural Engine, decides whether the request is local or needs escalation. The classifier is under 5 megabytes and runs in single digit milliseconds. Helpers handle weather, system metrics, network telemetry, log analysis, and trading. The Mac bridge is locked behind a shared secret, the request size is capped, and the label set is whitelisted. SmallIntent, the model, is mine. I trained it on 60+ labeled examples and quantized it to fit.

highlights

  • 01 On device classifier. No cloud round trip for routing.
  • 02 SmallIntent model. DistilBERT backbone, INT8 quantized, under 5 MB, ANE accelerated.
  • 03 Helpers are sandboxed. Schema validated, two step confirmation for critical actions.
  • 04 Bridge is hardened. Shared secret auth, request size caps, label whitelist, LAN exposure opt in.
  • 05 Multi domain. Weather, system, network, trading, DevOps, all from one voice interface.

stack

SwiftCoreMLPythonFastAPIPyTorchOllama

links

back to work