ClosetClear
An AI wardrobe app: photograph your clothes and it cuts out the background, tags them, and suggests weather-aware outfits — so getting dressed is faster and your closet stays minimal.
ClosetClear digitizes your wardrobe so it actually works for you. You photograph the clothes you own, and the app handles the rest: it cleans up the photo, tags the item, and learns what goes with what. Then it tells you what to wear based on the weather. I built it partly as a serious practice project — a real, production-shaped ML system end to end — and partly because I like the idea behind it: if you genuinely use what you already own, you buy less, waste less, and it's easier to keep a small, useful wardrobe instead of a cluttered one.
What it does
Snap a photo and the item gets added to your digital closet, automatically tagged and ready to match. When you ask it to style you, it factors in the current weather so the suggestion is actually wearable that day, not just visually nice. The everyday payoff is simple: less time deciding what to wear, a wardrobe you can see at a glance, and a nudge toward keeping things minimal and intentional.
How it works
The pipeline leans on a few models working together. U2-Net removes the background so each garment is isolated cleanly. Gemini 2.0 Flash reads the item and tags brand, model, and color. CLIP embeddings stored in a Pinecone vector index let it find pieces that are visually and stylistically compatible, and OpenWeatherMap feeds the live conditions into the "Style Me" recommender. The app is React Native / Expo with Tamagui for the UI, backed by Supabase for data, auth, and storage; the ML lives in a FastAPI service deployed on Hugging Face Spaces.
What I learned
The interesting challenge was making heavy ML feel snappy on a phone. I kept models resident in memory so they don't reload per request, offloaded the CPU-bound vision work off the main event loop, cached results by hashing the image so repeat work is instant, and used quantized ONNX inference for the embeddings. Getting all of that to feel instant — while staying within free-tier infrastructure — taught me more about real performance engineering than the model work itself did.