Building a Local LLM-as-Judge Pipeline for Image Dataset Curation

An AI-powered travel app revamped its photo accuracy by integrating a fast online resolver with a robust offline verification system. By employing local vision models like Gemma 3 and an iterative rule-based approach, the app significantly cut down incorrect image matches while keeping user-side delays minimal. This new method underscores the potential of local large language models to streamline and enhance data curation processes, offering a scalable solution for various industries dealing with large datasets.