Case studies
Four at a private-equity-owned national distributor — eight distribution centers, 67,000 stocked SKUs, an ERP nobody trusted and twenty buyers each running a private spreadsheet. One at a high-growth food and beverage brand with no plant of its own, where supply ran entirely through co-manufacturers. Company and vendor names are withheld. Every figure below is one I can walk you through.
The ERP suggested quantities from a three-month rolling average, which undershot every seasonal peak and then kept ordering after it passed. Buyers learned to ignore it. Replacing it meant building segmented safety stock, a workbench that surfaces exceptions instead of rows, and an AI agent that reviews the suggested PO lines nobody has time to read.
Sales, operations and finance each quoted a different fill rate, each defended it, and none of them was wrong — they were measuring different things and calling it the same word. Fixing the definition was the precondition for a SIOP process that could actually arbitrate anything. Then the cadence itself: split demand reviews converging into one enterprise supply review, with the overrides written down.
Two problems that look unrelated and aren't. Items that stock out stop being ordered, so the history records the shortfall as a drop in demand and the forecast learns to buy less of exactly what customers wanted most. Meanwhile the signals that actually move a seasonal category — degree days, fuel prices, residential permits — sat outside the plan entirely. Both were fixable with data the company already had access to.
At the food and beverage business: which carrier gets which freight lane, which co-manufacturer runs which SKU at what volume, and what should move between nodes instead of getting made. Each one is a mixed-integer program with constraints that bind against each other — capacity, minimum runs, commitments, service, shelf life. All three built and solved in Gurobi. Results described qualitatively here; numbers on a call.
Every function had its own version of the numbers and most of them were rebuilt by hand each week. The fix wasn't another dashboard — it was fixing the metric definitions, a conformed star schema underneath them, a two-layer architecture that survives concurrent users and can be rebuilt from a repository, and a funded team whose job is to own it.
A note on anonymity: I won't name a client, and I won't publish anything a competitor could use. What I will do on a call is walk you through the actual queries, the actual failure modes, and the parts that didn't work the first time — which is usually the more useful conversation anyway.
The systems behind the cases
The first three cases produced software that's still in use, and there are four walkthroughs running the real interfaces — synthetic data, vendor names genericized. The optimization models aren't demoable in a browser; that one you interrogate on a call.
Next step
These ran alongside full operating roles, over years, at two companies. As an engagement, the same work is scoped, sequenced and sized in a three-week diagnostic — and if the opportunity isn't there, that's what the readout says.