A production NLP system is three separable layers: a data layer that defines labels and splits, a model layer that adapts a pretrained checkpoint to those labels, and a serving layer that exposes predictions behind a stable interface. This capstone builds all three with the Hugging Face stack -- datasets, tokenizers, transformers, evaluate, PEFT, and the Hub -- on a synthetic customer-support ticket corpus you generate in step one, then stresses them with a label-drift change request that forces each layer to flex independently.
Skill Level: Beginner