DoorDash Feature Trails

Feature trailActive

AI-Powered Food Metadata Platform

DoorDash built an AI-led restaurant metadata platform that infers item- and store-level attributes using multimodal signals. Key innovations include an LLM jury system for high-quality evaluation, context-optimization agents for iterative prompt improvement, distributed computing for high-volume LLM inference, and AI-led annotation for training data generation. This platform enables reliable and cost-effective generative AI deployment at scale, improving the engineering workflow and consumer experience. The system now uses a reinforcement learning-inspired auto-context optimization loop for prompt development, AI-led data annotation for training data, and a distributed LLM inference pipeline with deduplication, Spark distribution, and batch processing for large-scale generation.

2 posts · 2026

Timeline

Feature trailActive

Search Relevance Prediction

DoorDash implemented a new relevance prediction system for search ads, leveraging small language models (SLMs) to improve the accuracy of ad matching to user search intent. This system uses an LLM for offline label generation and an SLM for low-latency online prediction, employing a bi-encoder architecture with BERT-based encoders and 64-dimensional embeddings. The system now uses a DistilBERT bi-encoder with shared encoder weights, CLS pooling, 64-dimensional embeddings, and an online bilinear layer for low-latency relevance scoring, trained using CORAL for ordinal regression.

1 post · 2026

Timeline