DoorDash Uses LLM Juries to Build Food Metadata

DoorDash’s engineering team announced a new system for generating detailed food metadata that leverages large language models (LLMs) and multimodal AI. The approach, described in a recent blog post on the company’s careers site, combines text‑based LLMs with image‑recognition models to extract ingredients, nutritional facts, and allergen information from restaurant menus. By employing a “LLM jury” framework—where multiple models evaluate and vote on candidate outputs—the team improves the accuracy and consistency of the data, while a context‑optimization layer tailors prompts to the specific culinary domain. The solution replaces labor‑intensive manual tagging, enabling DoorDash to scale its menu catalog across thousands of partners more efficiently.

The initiative, which earned 13 points on Hacker News, reflects a broader industry push to enhance the reliability of food‑delivery platforms through AI‑driven automation. Accurate metadata supports better search, personalized recommendations, and compliance with dietary regulations, ultimately improving the customer experience. DoorDash plans to continue refining the system, extending its multimodal capabilities and integrating feedback loops to maintain data quality as the platform expands.