Automate replenishment with MMF, Databricks Genie, and Amazon Quick
aws ·
This AWS Machine Learning Blog post is a technical walkthrough for building an unattended retail replenishment loop using Databricks Many Model Forecasting (MMF), Databricks Genie, and Amazon Quick. It argues that because foundation models can now forecast demand across an entire catalog without per-item tuning, the bottleneck has shifted to the gap between a good forecast and an order placed on it, since forecasts, supplier availability, and the person placing orders live in different systems. The proposed solution runs in four stages: Databricks MMF serves Chronos-2 to predict 7-day demand for every SKU; a Databricks Genie Agent detects surging SKUs, defined as a next-7-day average at least 1.5 times the prior-14-day average and with a prior-14-day average of at least 1; Amazon Quick reconciles each surging SKU against live supplier availability in Amazon S3 Tables and chooses the cheapest supplier that can cover it; and Amazon Quick Flows either places a routine purchase order through a Supplier Order API or raises a human-review ticket when no single supplier can cover the surge. Amazon Quick sits in the middle as the only component touching both worlds, reaching the forecast through a Databricks Genie Agent over the Model Context Protocol, pulling the supplier feed from S3 Tables, and calling the Order API through an OpenAPI connector, joining them on a shared product key. The post covers prerequisites, environment setup, two automation scripts (setup_databricks.sh and setup_aws.sh), several console-only steps, and the OAuth and role requirements for Amazon Quick and Databricks. It also describes a synthetic supplier feed derived from FreshRetailNet-50K product keys that is deterministic at 63,861 rows, with one supplier set deliberately under-stocked to exercise the human-review path, and notes that Amazon Quick Enterprise edition is the main recurring cost.
aws