Objective
Move from keyword matching to lightweight intent understanding for spoken reminders and commitments.
Issue
Simple keyword rules produce too many false positives and feel rigid. Full LLM calls on every utterance are too slow and power-hungry for a desk device.
Observation
A small set of high-frequency intents (remind me, I will finish X by Y, cancel previous, status check) covers most real usage so far.
Finding
Hybrid approach works: local keyword + short pattern matching for common intents, and a deferred LLM call only when the pattern is ambiguous or new.
Next Step
Implement the hybrid classifier on-device and log real utterances for later refinement. Keep the LLM path optional and offline-first.