Automatic Labeling Model

The Operational Bottleneck KakaoStyle, a leading South Korean e-commerce platform, launched a new feature allowing users to filter reviews by specific clothing attributes like "Fit" and "Length." However, training the underlying model required thousands of labeled reviews. The Data Science team was bogged down by manual data labeling, creating an inefficiency in the product development lifecycle. The Automated Solution I identified this workflow gap and engineered an automated labeling pipeline to replace the manual process. I fine-tuned a BERT model on internal review datasets, training it to classify both general sentiment (Positive/Neutral/Negative) and domain-specific attributes (e.g., "True to size"). Deployment & Tooling To make this accessible, I wrapped the model in a web interface using Streamlit. This allowed stakeholders to simply upload a raw CSV and receive a fully labeled dataset. The tool transformed the workflow from manual tagging to rapid verification.