# SoundSense AI - Intent Classifier

This is a fine-tuned model checkpoint for SoundSense AI.

Base model / source:
- bert-base-multilingual-cased, https://huggingface.co/bert-base-multilingual-cased

Training data:
- Custom smart-home command dataset (1050 training / 200 validation samples) 
  covering 10 intent classes: check_weather, control_fan, lock_door, play_music, 
  set_alarm, set_temperature, stop_music, turn_off_light, turn_on_light, unlock_door

Use:
- Part of SoundSense AI hackathon submission (Stage 7: Intent Classification & Routing).
- Classifies transcribed ASR text into a smart-home intent and routes to a 
  personalized per-user device response.

Limitations:
- Built for prototype/demo use.
- Achieved val F1 = 1.0 at epoch 6, but on a small, clean, curated dataset 
  (200 val samples); real-world generalization to noisy/Hinglish ASR transcripts 
  not yet verified.
- Performance should be verified on the target environment before deployment.