For Waze, accurate road closure data is critical. Minor errors can cause failed routes and erode user trust. Much of this data comes from public sources. However, the highly flexible and diverse nature of these data sources creates a severe bottleneck, making it difficult to scale via traditional methods. To overcome this, we deployed a GenAI powered pipeline, leveraging few-shot in-context learning to autonomously translate data into standardized spatial objects. This talk unpacks how we shrunk feed onboarding to minutes. We will explore the architecture behind this scale-up, detailing our rigorous Human-in-the-Loop (HITL) rollout, and offer practical lessons on building decoupled 'LLM-as-a-Judge' evaluation suites to monitor online quality and accelerate iterative model optimization.
Senior Data Scientist, Waze (Google)