Noma | Devpost

Front Page
Questions for the User
Home Page
Results

Inspiration

Skin cancer is the most common cancer worldwide, yet early detection increases survival rates to 99%. The problem: dermatologist access is limited, especially in underserved areas. We wanted to put a preliminary screening tool in everyone's pocket.

What it does

Noma analyzes skin lesions using AI to provide informational insights.

Questionnaire — Location, duration, and recent changes (like a dermatologist would ask) Camera — Guided capture with crosshair overlay Analysis — AI classifies the image and generates a contextual summary Results — Confidence score, color-coded severity, plain-language summary, and next steps

Always emphasizes: this is not a medical diagnosis.

How we built it

React Native + Expo — Cross-platform mobile app HuggingFace — Skin lesion classification model (trained on 50,000+ images) Google Gemini — Combines image results with questionnaire answers for contextual analysis react-native-reanimated* — Smooth animations **react-native-svg — Animated progress indicator

Challenges we ran into

When we were creating the prompt to Gemini that analyzes our data, we had a challenge crafting responsible messaging that informs without alarming.
Handling variable image quality from different devices / lighting / zooming (in/out)
Version control (merge conflicts)

Accomplishments that we're proud of

Polished, calming UX appropriate for a health app
Smooth connection of AI model / pipeline (Gemini) to analyze hugging face model and user data
Responsible design with clear disclaimers throughout

What we learned

Integrating multiple AI services (Hugging Face + Gemini)
Prompt engineering for structured LLM outputs
Mobile image manipulation and optimization
Designing health apps that balance helpfulness with responsibility

What's next for Noma

User accounts to track changes and provide more informative results
Try to connect with local dermatologists (telehealth)
Clinical validation and expand with more models for expanded condition condition detection