DermAI Pakistan
On-device dermatology decision support for low-resource clinics. A quantized deep learning classifier that runs entirely on the phone.
The problem
Dermatology expertise is scarce outside major cities, and connectivity in rural clinics is unreliable. DermAI brings a decision-support classifier to the clinician's own device, so inference never waits on a server round trip.
Approach
- EfficientNetV2-S classifier trained for common skin conditions, then quantized to INT8 TFLite for fast, low-memory inference on mid-range Android hardware.
- Native C / ARM NEON preprocessing bridged into Flutter through Dart FFI, so the image pipeline runs at native speed rather than in the Dart VM.
- Explainability stays server-side, where Grad-CAM and SHAP produce heatmaps and attributions for clinician review without bloating the on-device model.
Why it matters
The whole inference path works offline and degrades gracefully where bandwidth does not exist. A cloud layer adds interpretability whenever a connection is available.
More work
View all
ml_frame
A zero-dependency machine learning framework in roughly 2,800 lines of C11: arena allocator, reverse-mode autograd, multi-backend GEMM, and INT8 quantization.
Clinic Appointment Booking
An embeddable AI assistant that answers patient questions and books appointments from any clinic website. Inventix's flagship product.