Building Food Metadata with LLM Juries, Context Optimization & Multimodal AI - DoorDash

This post details the technical innovations in DoorDash's AI-led food metadata platform, focusing on an LLM jury system for evaluation (increasing accuracy by ~20%), a reinforcement learning-inspired auto-context optimization loop for prompt development (increasing speed tenfold and precision by >20%), AI-led data annotation for training specialized fine-tuned models (reducing inference costs by ~90%), and a distributed LLM inference pipeline using deduplication, Spark, and batch processing to cut backfill time from over a month to days.