Anujkudu/VaniCure-

Offline AI-powered respiratory screening tool that analyzes cough/breathing audio to flag early signs of TB and Asthma — built for low-resource clinics. # VaniCure **Offline AI-powered respiratory screening for low-resource clinics.** VaniCure listens to a patient's cough or breathing and flags potential early signs of **Tuberculosis** and **Asthma** — entirely on-device, with no audio ever leaving the clinic. It's built for health workers in rural or low-connectivity settings who need a fast, privacy-safe first pass before referring a patient onward. --- ## Why VaniCure exists Respiratory screening tools are often cloud-dependent, expensive, or require lab infrastructure that just isn't available in many rural clinics. VaniCure runs its entire inference pipeline locally — no internet required after setup — so a health worker with a laptop and a microphone can get a risk signal in seconds. ## What it does - **Records or accepts an uploaded audio clip** (cough, breathing) from a patient - **Runs three independent AI models** on that audio and combines their outputs - **Produces a risk score** for TB, Asthma, or Normal, plus a recommended next step - **Asks short multilingual follow-up questions** (Hindi, Gujarati, Marathi, English) to add context - **Stores everything locally** — patient history, scores, exportable as CSV - **Watches for clusters** of high-risk cases and raises an outbreak alert if it sees one ## The model pipeline | Model | Role | Notes | |---|---|---| | **PANNs (CNN14)** | Deep, general-purpose audio feature extractor | Trained on AudioSet, repurposed here for respiratory cues | | **YAMNet** | Lightweight, fast first-pass classifier | MobileNet-based, runs almost instantly | | **CNN-BiLSTM (custom)** | Purpose-built respiratory classifier | CNN for spatial features + BiLSTM for temporal patterns, >0.87 F1 on test data | Each model's output is blended into a weighted risk score (cough, wheeze, breathing, and throat-clearing probabilities feed into separate TB and Asthma formulas), then averaged across all three for a final verdict. ## Stack **Frontend** — React 18, TypeScript, Vite, Tail …