TexTell — Multimodal Fashion Tagging & Listing

A mobile app that turns a photo of a garment into structured attributes and a ready-to-post shop listing — backed by two models I trained on Kaggle and exported for inference: a PyTorch vision tagger and a LoRA-fine-tuned text generator.

The problem

Small fashion sellers lose hours doing the same two chores for every item: figuring out what to call it (category, material, style attributes) and writing a listing that will actually sell. TexTell collapses that into one step — photograph the garment, get structured tags and a drafted listing back.

Two models, one product

TexTell is genuinely multimodal, and each half is a real trained model rather than an API call:

a computer-vision tagger (PyTorch, ~24 MB exported) that reads the image and predicts garment attributes;
an NLP model fine-tuned with LoRA (a compact adapter, ~9 MB) that turns those attributes into listing text.

LoRA was a deliberate choice: parameter-efficient fine-tuning gets a domain-adapted generator without the cost or footprint of tuning a full model — which matters when the artifact has to ship inside a phone app.

Architecture, honestly

I trained both models on Kaggle and exported the weights; the repository is the productionization side — a FastAPI inference service that loads the exported .pt and LoRA adapter, and a React Native / Expo app with WatermelonDB for offline-first local storage. So the code you can read is the serving and app layer; the training notebooks live on Kaggle. I’d rather scope that precisely than imply the whole stack sits in one repo.

Reflection

TexTell is my clearest example of taking ML the last mile — not stopping at a notebook metric, but exporting the model, wrapping it in an inference service, and putting it behind a real mobile UX. The gap between “it works in the notebook” and “it works in someone’s hand” is where most of the engineering actually is.