All Work
AI / ML
NeuraGallery
AI-powered image clustering platform using DINOv3 vision transformer
Personal Project·2025
95%
Cost Reduction
Live
Status
3-tier
Architecture
Problem
Organizing large image collections manually is tedious. Existing solutions require manual tagging or expensive GPU instances for AI-based clustering.
Approach
Built a 3-tier microservices architecture: Next.js PWA frontend, FastAPI backend with Azure Blob Storage, and a PyTorch ML service using Meta's DINOv3 vision transformer. The key innovation is an embedding cache - compute embeddings once, then re-cluster in milliseconds without GPU.
Results
Live platform achieving 95% cost reduction vs GPU instances (~$30/month vs ~$550/month). Users upload photos and the system automatically groups them into coherent clusters using K-Means and HDBSCAN.
Technologies
PyTorchDINOv3FastAPINext.jsTypeScriptAzure Blob StorageDockerMongoDBRedisK-MeansHDBSCAN