Clarion
Offline AI-powered multimodal alert system that notifies deaf students of sounds they cannot hear.
- Trained an audio event classifier to a 91.5% weighted F1-score on 3,435 validated data points, deployed on an ESP32-S3 with a React Native companion app, running fully offline with no cloud dependency.
- Compressed a 300 MB model to 80.7 KB INT8 — a 99.97% reduction — through knowledge distillation, with quantitative analysis of the accuracy trade-off at each step.
- Awarded a Distinction on dissertation Chapters 1–3 covering problem definition, literature review and system design.
- Designed the multimodal alert pathway around deaf students' real constraints in Zimbabwean classrooms, where reliable connectivity cannot be assumed.
- Weighted F1
- 91.5%
- Validation set
- 3,435 data points
- Model size
- 80.7 KB INT8
- Compression
- 99.97% vs 300 MB
- Hardware
- ESP32-S3
- Companion app
- React Native
- Stack
- TensorFlow, Keras, TFLite
- Assessment
- Distinction, Ch 1–3
- Supervisor
- Mr G. Jekese, UZ

