A second-year BSc AI team project (Tweedejaarsproject) at the University of Amsterdam for the wine startup Vyno: an NLP system that turns dense sommelier wine descriptions into short, readable descriptions for everyday consumers.

Try it live (runs in your browser)

The actual fine-tuned T5 model, quantized to int8 and running fully client-side via transformers.js (ONNX Runtime Web), with no server and no API. First run downloads ~275 MB of weights (cached afterwards), then generates word by word right on your device.

Examples:
Model not loaded yet. Click Generate to download and run it.

Problem

Vyno makes wine more accessible to everyday consumers through Naomi, a virtual sommelier that recommends wines by asking a few simple questions. But sommelier descriptions are long, thorough, and full of domain-specific vocabulary, which makes wine feel inaccessible. These descriptions need to be shortened and simplified while staying informative, so consumers can still benefit from expert knowledge.

Approach

The task is Natural Language Generation (NLG): produce concise, descriptive text from an input of keywords. The pipeline extracts keywords from an existing sommelier description (or, when a wine has no description, from descriptions of the same grape and region), then a model generates a novel simplified description from those keywords. The model is trained on English descriptions, with multilingual support noted as future work so Naomi could serve non-English markets.

Original project demo video.

Materials