EasyDerma | Devpost

Inspiration

Growing up, many of our members had difficulty managing their acne breakouts especially with the amount of misinformation surrounding the topic online. Influencers promoting different products every day and discrediting perfectly fine alternatives for marketing purposes made it impossible to trust any online source.

In retrospect, having acne severely impacted our self image and overall proactiveness in our community. It has long been considered a natural and unavoidable part of adolescence, but we challenge the core of that established narrative.

What it does

Upload a picture of your face. Then, our personal AI model that we trained during the hackathon will take a look and diagnose some conditions, then Gemini API will provide a second opinion. Based on the agreed conditions, they will provide the user with a morning and night skincare routine, along with hyperlinks to amazon to purchase them. They will also be provided with instructions on how and why to use these products, and, finally, some general dietary recommendations.

How we built it

The UI was vibe coded in Next.js, then added Auth0. The backend was made with Flask and communicates with MongoDB, Gemini API, and uses our convolutional neural network. We built our CNN with a Kaggle dataset consisting of 27 000 images of skin diseases. Using tools like TensorFlow's built-in Keras model, we applied 6 layers of convolution, max pooling, and data augmentation to maximize the extraction of relative features and avoid overfitting.

Challenges we ran into

Vibe debugging, uncooperative dataset, compute limitations, bad AI results, storing images in a database.

Accomplishments that we're proud of

We are most proud of our beautiful UI, the practical benefits that EasyDerma provides, making everything in sync across devices with MongoDB, training our own AI model.

What we learned

Some of the things we learned during this hackathon are the reality of applied machine learning, vibe coding, vibe debugging, the Next.js framework, and Auth0.