Simplify
A web app that identifies likely skin conditions from a photo and tells you what you can
do about it right now — immediate care, home remedies, what to avoid, and the red flags
that mean you should stop reading and see a doctor.
Classification runs entirely in the browser via a Teachable Machine model, so the photo
never leaves the user's device.
Live: simplify-skin-disease.vercel.app
This is not a medical device. It is an image classifier that can be confidently
wrong. Every result is a prompt to get checked — never a reason to delay care or to
self-treat something serious.
Screens
Home
Skin Guide
Reference material on the conditions the model is trained to recognise.
About
Log in
Sign up
Password requirements tick off live as you type.
Prediction
Drag-and-drop or webcam capture, with distinct model-loading and scanning states.
Result and guidance
Shown here with sample data to illustrate the layout.
Features
- Email/password auth via Firebase, with sessions that persist for 7 days
- In-browser classification — no image upload, no server-side inference
- Upload or webcam — webcam predicts on a frame you capture, not continuously
- Condition guidance for 20 conditions: immediate care, home remedies, what to
avoid, and when to seek medical help
- Severity routing — anything cancer-adjacent skips home remedies entirely and
directs to urgent care
- Confidence reporting, with an explicit warning below 60%
Model
A Teachable Machine image model with seven
classes:
| Class | Guidance shown |
|---|
| Acne and Rosacea | Acne & Rosacea |
| Molluscum contagiosum and warts | Molluscum Contagiosum / Viral Skin Infection |
| Melanoma | Possible skin cancer — urgent |
| Vascular Tumor | Vascular Tumour |
| Eczema | Eczema (Atopic Dermatitis) |
| Bullous | Bullous Disease — urgent |
| Psoriasis and lichen | Psoriasis |
Labels are matched loosely against static/assets/js/remedies.js, so variants like
"Atopic Dermatitis" or "Tinea Corporis" resolve correctly. An unrecognised label
falls back to a generic card rather than breaking the page.
Stack
| Layer | Choice |
|---|
| Backend | Flask, deployed as a Vercel serverless function |
| Auth | Firebase Authentication (REST API) |
| Model | Teachable Machine + TensorFlow.js, client-side |
| Frontend | Jinja templates, vanilla JS, no build step |
Layout
api/index.py Flask app — all routes, the serverless entry point
templates/ Jinja templates
static/assets/js/
skin-scan.js Upload, webcam capture, model loading, result rendering
remedies.js Condition guidance and label matching
auth.js Password toggle, live rules, submit state
static/assets/css/
auth.css Login and signup
vercel.json Serverless build and routing
app.py in the repo root is the original single-file version, kept for local
reference. Vercel serves api/index.py.
Running locally
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export SECRET_KEY="$(python3 -c 'import secrets; print(secrets.token_hex(32))')"
python api/index.py
Then open http://127.0.0.1:5000.
Deploying
Pushing to main triggers a Vercel build. vercel.json routes everything to the
Python function except /static/*, which is served from the CDN, and includeFiles
ensures the templates are bundled into the function.
Environment variables
| Name | Required | Purpose |
|---|
SECRET_KEY | Yes, in production | Signs session cookies. Without it the app falls back to a default that is committed to this repo, which means anyone could forge a session. Generate one with python3 -c "import secrets; print(secrets.token_hex(32))". |
The Firebase Web API key in api/index.py is intentionally public — that is how Firebase
web keys work. Access is controlled by Firebase security rules, not by hiding the key.
Known limitations
/map renders map.html, which does not exist in templates/ — the route returns a
500. Nothing links to it.- Model accuracy is bounded by its training set. Treat low-confidence results as noise.
Licence
MIT — see LICENSE.
