Auto-tag fashion products with ResNet50 + XGBoost

01 Hook

Manual product tagging doesn't scale when every SKU needs color, type, and a price guess.

02 Context

End-to-end ML product: transfer-learned ResNet50 for vision, XGBoost for price, Flask inference API, React (Vite) upload UI.

03 Problem

E-commerce catalogs burn time on repetitive labeling. Separate vision and pricing models need one coherent product surface.

04 Insight

Freeze most of ImageNet ResNet50 and train a thin head — get strong tags without training a CNN from scratch.

05 Exploration

Multi-label color/type outputs (10 + 10), Adam + dropout regularization, and a second tabular model for price from metadata.

06 Decision

Kept the system stateless (no DB) so inference stays deployable as a simple backend/models bundle.

07 Solution

Shipped upload → classify color/type → predict price via REST, with documented model files and local run instructions.

08 Impact

Proxy proof: 20-class vision head, ~24.8M-param ResNet50 transfer setup, dual-model architecture in a public repo recruiters can open.