Agentic Machine Learning Modeling at Instacart

This post details the application of agent-assisted modeling loops to two specific Instacart use cases: Delivery Time prediction and Catalog Attribute Extraction. For Delivery Time, agentic autoresearch loops achieved 3-5% offline MAE improvements over a mature gradient-boosted tree baseline by exploring different model families (LightGBM, MLP), tuning hyperparameters, and combining smaller improvements like recency weighting and seed averaging. For Catalog Attribute Extraction, an agent-driven optimization loop reduced manual optimization time from weeks to days, achieving significant recall gains (up to 9.1 points) while maintaining precision by adapting model choice (GPT-5.4-mini, o4-mini) and refining prompt instructions based on error analysis. The post highlights the adaptive nature of these loops, where agents learn from failed trials and revise approaches, and discusses the infrastructure reliance on Griffin and ongoing development of specialized agentic modeling infrastructure.