DoorDash serves a vast and diverse set of merchants, with every restaurant, menu, and dish expressed in its own unique way. A high‑quality food catalog forms the backbone for customer search and personalization, representing a key driver of restaurant success. Unlike standardized catalogs, food is deeply contextual, culturally rich, and highly non‑standardized; the same dish can be described in countless ways, while entirely different dishes may share similar names, descriptions, or images. Add to this that, at DoorDash’s scale, there are millions of unique items and constant menu updates. This variability and volume make it extremely challenging to generate reliable metadata via traditional approaches.

To address this, we built an AI‑led restaurant metadata platform. Our platform infers item‑ and store‑level attributes — for example, whether an item is spicy, or that a restaurant’s cuisine is Chinese — using multimodal signals from text, images, and broad web searches. To build trustworthy, accurate metadata at scale, we engineered several key innovations within the complex DoorDash system, including: a LLM jury system for high‑quality evaluation, which increased the annotation accuracy by roughly 20% compared with typical human reviewers; context‑optimization agents to iteratively improve prompts within minutes, increasing model precision by more than 20% while avoiding the inefficiency of handcrafted, suboptimal prompts; this loop accelerated prompt development tenfold; distributed computing enables high‑volume LLM inference, cutting backfill time from over a month to just a few days; and AI‑led annotation to generate training data, which unblocked fine‑tuning to match frontier LLM quality at 10% of the inference cost, with zero human annotation effort.

The flow begins with ingesting menu updates and deduplicating to minimize inference costs. Items are fed into AI generators to produce metadata, which undergoes immediate structural validation for error detection and retries. We continuously monitor the quality of generated predictions via an LLM jury; the evaluation result is also used for context engineering to improve generation quality. Additionally, we provide a merchant override mechanism that allows business owners to validate or correct attributes. This metadata platform allows us to successfully deploy generative AI reliably and cost‑effectively at scale, improving our engineering workflow and the DoorDash consumer experience.