The Problem
Vision-language fashion recommender that learns a shopper's taste from just a handful of interactions.
VISION-LANGUAGE · RECOMMENDERS
StyleNova
StyleNova fuses what an outfit looks like, what a shopper has liked before, and what similar shoppers bought into a single hybrid recommendation score. CLIP vision-language embeddings, TF-IDF content signals, and collaborative filtering feed the base model, while an online preference-learning loop keeps adjusting to each shopper in real time.
GitHub · soonLive demo · soon
Stack
AI + ML
PyTorchCLIPscikit-learn
Backend + Data
PythonFastAPIPrisma
Frontend
Next.jsTypeScriptZustandFramer Motion
What I Built
StyleNova fuses what an outfit looks like, what a shopper has liked before, and what similar shoppers bought into a single hybrid recommendation score. CLIP vision-language embeddings,…
The Difference
Hybrid scorer (CLIP ViT-B/32 + TF-IDF + collaborative filtering) reaches Precision@10 of 0.73
Numbers That Matter
32primary scale
10measured result
0.73system signal
32%quality outcome
Engineering Notes ✎
Engineering Note 1
Hybrid scorer (CLIP ViT-B/32 + TF-IDF + collaborative filtering) reaches Precision@10 of 0.73
Engineering Note 2
Online preference learning improves recommendation relevance 32% after just 5 user interactions
Engineering Note 3
Cached CLIP embeddings deliver sub-200ms p95 API response for 10K+ concurrent users with zero cold-start
Architecture Overview
Taste signals↓Hybrid Recommender
CLIP EmbeddingsTF-IDF ScorerPreference Loop
Prisma Store↓Personalized looks