An AI that knows when it doesn't know.
A hybrid audio model that classifies infant cries and, crucially, refuses to answer when it shouldn't. Built for a healthcare context where a wrong 'confident' answer is worse than 'I don't know'.
Standard cry classifiers optimize accuracy on curated datasets and ship as black-box baby-monitor features. In deployment, ambient noise, unfamiliar voices, and rare cry types produce over-confident, misleading predictions.
The model surfaces plain-English uncertainty reasons instead of hiding low-confidence guesses, which is the whole point of trustworthy AI in a healthcare-adjacent setting.
Trained a CNN + CRNN (BiLSTM) on mel-spectrograms across pain/hunger/discomfort/tired/other classes with W&B experiment tracking.
Added an uncertainty layer combining softmax entropy, top-1 margin, and OOD detection — flagging predictions the model shouldn't stand behind.
Wrapped everything in a Streamlit dashboard with spectrogram visualization so caregivers see why the model made a call.
Deployed a FastAPI inference service behind React on Render for a smooth demo experience.
"Being loud about 'I don't know' was more useful than another 2% on the confusion matrix."