Description

Lumeskin is a cutting-edge web platform that leverages advanced AI to provide accessible and instant skin analysis. Users can upload their images, get real-time computer vision-powered diagnosis, and explore detailed information about detected conditions. The monorepo contains all code and assets: ML models (training and scripts), REST API backend (FastAPI), and the complete Next.js frontend

How I Made It

• Sourced diverse dermatology image datasets from Kaggle and other open medical sources for robust training data.
• Trained and fine-tuned advanced computer vision models (YOLO, ResNet, and custom CNNs) on Google Colab using PyTorch, with extensive experimentation on batch size, learning rates, and augmentations.
• Achieved an approximate 70% detection rate for skin lesion/classification after multiple epochs and training runs, mainly limited by the number and variability of available datasets.
• Developed the backend REST API using FastAPI (Python), enabling real-time ML model inference and seamless communication between frontend and ML logic.
• Built the frontend as a fully responsive Next.js application (React 18), designed for fast uploads, instant feedback, and mobile usability.
• Used Framer Motion for smooth page transitions and interactive UI animations.
• Integrated advanced error handling, input validation, and user feedback for uploads and predictions.
• Collaborated using GitHub monorepo for tight integration of ML, design, backend, and frontend assets.
• Created detailed Figma flows and branding guidelines in the /design repo directory.

Challenges Faced

• Normalizing image submissions for consistent inference accuracy across devices, lighting, and skin tones.
• Managing secure ML inference hosting to deliver low-latency results at scale.
• Ensuring a privacy-conscious approach for sensitive health/data uploads, including GDPR best practices.
• Bridging frequent updates to the ML model and seamless deployment to production.

Key Wins

• Shipped a fully functional, public AI skin diagnostics app.
• The real 'star' is our custom-trained PyTorch (.pt) model, which approaches 70% detection on challenging, real-world skin images – the main limiting factor being the size/diversity of the available datasets.
• Built an integrated product from model training → API → frontend deployment in a single repo.
• Provided instantly actionable information to users within seconds of uploading an image.
• Enabled non-experts to get accessible, high-confidence preliminary skin health insights through AI.
• Established open, documented architecture for further community or clinical extension.