DermAI

Inspiration

Millions of people skip dermatologist visits because of cost, long wait times, and lack of access. We wanted to build a first-opinion skin analysis tool that anyone with a smartphone could use — no appointment, no insurance, no barrier.

What it does

DermAI lets users upload a photo of a skin condition and instantly receive an AI-powered analysis identifying conditions like acne, rashes, moles, hair loss, and nail problems. Each result includes a confidence score, severity rating, personalized advice, and an AI chat interface for follow-up questions.

How we built it

ML: Trained an EfficientNet-B0 model on Google's SCIN dataset across 8 skin condition classes, then exported to ONNX for faster inference Backend: Flask REST API with a PyTorch + ONNX ensemble, JWT authentication, SQLite scan history, and a DeepSeek-powered diagnosis chat Frontend: React + Vite + Tailwind with image upload, camera capture, and a clean clinical chat interface

Challenges we ran into

The SCIN dataset has significant class imbalance — RASH dominates — which required careful augmentation strategies and stratified splits to prevent the model from over-predicting common classes. Getting consistent predictions across the PyTorch and ONNX ensemble was trickier than expected, and calibrating the AI chat to be medically responsible without being so restrictive it becomes useless was a real balancing act.

Accomplishments that we're proud of

We built a full end-to-end pipeline — from raw SCIN data on HuggingFace all the way to a working web app — in a single hackathon session. The multi-model ensemble with calibrated confidence scoring is something we're particularly proud of, along with a UI clean enough that non-technical users can actually navigate it without guidance.

What we learned

Medical AI demands a different standard of care than typical ML projects — confidence communication and disclaimers aren't just legal cover, they're part of the UX. We also learned that ONNX export meaningfully speeds up inference with minimal accuracy loss, and that solid data preprocessing consistently matters more than chasing a fancier model architecture.

What's next for DermAI

Expand to more skin condition classes with better coverage of diverse skin tones Add scan history trends so users can track changes over time Build a mobile app with real-time camera analysis Partner with telemedicine platforms to enable seamless doctor referrals when the model flags high-severity cases

Built With

• deepseek-api
• efficientnet-b0
• flask
• google-scin-dataset
• jwt
• onnx-runtime
• python
• pytorch
• react
• scikit-learn
• sqlite
• tailwind-css
• timm
• vite