Automate replenishment with MMF, Databricks Genie, and Amazon Quick
Amazon Web Services Venkatavaradhan Viswanathan
AWS shows a retail replenishment loop that spots surges, picks a supplier, and can place the order on its own. It stitches Databricks forecasts to Amazon Quick, so the forecast finally turns into action.
Based on reporting by Amazon Web Services, Venkatavaradhan Viswanathan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS’s pitch here is simple: forecasts are only useful if something can do the buying. In retail, that handoff is usually where things fall apart. A planner exports a forecast, checks supplier availability, then tries to work through tens of thousands of SKUs while the fastest movers are already running dry. The post argues that foundation models solved the first half of the problem, and that the real bottleneck is now the gap between prediction and purchase.
The build splits that job across Databricks and Amazon Quick. Databricks runs Many Model Forecasting with Chronos-2 to predict 7-day demand for every SKU. A Databricks Genie Agent then flags surges using a clear rule: the next-7-day average must be at least 1.5 times the prior-14-day average, and that prior-14-day average must be at least 1. From there, Amazon Quick takes over, reading live supplier availability from Amazon S3 Tables, choosing the cheapest supplier that can cover the surge, and either placing a routine order or escalating to a human when no single supplier fits.
The interesting bit is how little gets copied around. Quick sits in the middle, talking to the Genie Agent through the Model Context Protocol, reading the supplier feed through a Direct Query dataset, and calling the Supplier Order API through an OpenAPI connector. Everything joins on a shared retailer_product_id at decision time. Databricks supplies the forecast; Quick turns it into an action.
The setup is designed for a real terminal, not a slide deck. AWS lists the Databricks CLI 0.299.0+, AWS CLI 2.36.2+, jq 1.7, and uv or Python 3.11 as the main tools. The repo includes scripts for the Databricks side and the AWS side, plus a few console-only steps for the action connectors, the S3 Tables access grant, and the dataset. On the Databricks side, notebooks 01 and 02 come from the fresh_retail_net accelerator, while notebooks 04 and 03 are shipped in this repo to expose the forecast through Genie.
On the AWS side, the supplier feed is synthetic but deterministic: the loader generates three suppliers per product and city over a 7-day window, and every run produces the same 63,861-row feed. One supplier set is intentionally under-stocked so the human-review path actually gets exercised. The Supplier Order API is just a stand-in, built with API Gateway, Lambda, and DynamoDB, and the flow writes either orders or tickets through it. The whole thing is framed as a demo, but the architecture is clearly aimed at a familiar enterprise problem: forecasts are easy to admire and annoyingly hard to act on.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of automation: boring, specific, and slightly ruthless about where the human belongs. The industry has spent years worshipping forecasts like they were the finish line; they’re not, they’re the first sentence. Also, a system that only interrupts people when no rule fits is a lot more honest than the usual “AI copilot” fluff with a shiny button attached.
Read more about this at: Amazon Web Services