Create the perfect blend of data to fix class bias in computer vision datasets before fine-tuning models.

PerfectBlend inspects a COCO-format dataset, identifies class imbalances with with AI-powered, context-aware analysis, and rebalances the dataset through downsampling and data augmentation.

• Connect a local COCO dataset by directory path
• Visualize per-image class distribution as an interactive packed bubble chart using D3.js
• Analyze class bias using Gemini + RailTracks, grounded in real-world domain citations
• Adjust AI-suggested annotation targets before applying any chances so that you stay in control
• Rebalance via downsampling majority classes and augmenting minority classes; visualize and view the new exported balanced dataset

Frontend:   React · TypeScript · Tailwind CSS · D3.js
Backend:    Python · FastAPI · RailTracks · Gemini API · Albumentations

Backend: run from the api/ directory:

cd api
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env          # then fill in your API keys
uvicorn main:app --reload

Frontend: run from the project root:

npm install
npm run dev

Create api/.env and fill in:

GEMINI_API_KEY=your_key
GEMINI_MODEL=your_model      # e.g. gemini-3-flash

skin_dataset/ is included — a pre-split COCO dataset with train/valid/test splits ready to use. Point the app at the absolute path to skin_dataset/ to walk try out PerfectBlend.

Your dataset must follow this structure:

your_dataset/
├── train/
│   ├── _annotations.coco.json
│   └── images (*.jpg / *.png / *.webp)
├── valid/
│   ├── _annotations.coco.json
│   └── images
└── test/
    ├── _annotations.coco.json
    └── images

Note: only the train/ folder will be rebalanced.