Challenge
Ferrum Logistics ran nine warehouses across India and Southeast Asia, each with its own inventory system, its own purchasing rhythm, and its own informal spreadsheet for "the stuff we always run out of." Replenishment decisions were made weekly, by hand, by warehouse managers who relied on memory, last-quarter's data, and the occasional WhatsApp message from a regional director.
The result was the worst of both worlds: chronic stockouts on fast-moving SKUs, and quietly bloated inventory of slow-movers that nobody wanted to write down. An internal audit estimated $1.2M in annual revenue lost to stockouts and another $4M tied up in stock that turned less than once a year.
The project started, as these often do, with a request for "a dashboard." It became something more interesting once we mapped the actual decision-making process.
Approach
The first deliverable was not software. It was a memo. We spent three weeks shadowing four warehouse managers, mapping the decisions they made each week and where the information came from. The memo identified six recurring decisions, ranked them by leverage, and recommended automating two — replenishment and slow-mover liquidation — while leaving the others to humans.
The technical build was deliberately small. A Postgres warehouse aggregating inventory feeds from each location, a Python forecasting service running on Lambda for daily SKU-level predictions, and a Next.js operations console that surfaced recommended orders to the warehouse managers — but did not place them automatically. Every recommendation came with a confidence score and a one-paragraph rationale.
We made one architectural choice that paid off repeatedly: every recommendation was logged with the data that produced it. When a manager overrode the system, we captured the override and the reason. Six months in, those overrides have become the most valuable training signal we have.
The humans stayed in the loop. They click "approve" on most recommendations, edit a few, and reject the rest. The system handles the volume; the people handle the judgment.
Outcome
Six months after rollout, stockouts are down 71%, eliminating an estimated $1.2M in lost annual revenue. Slow-mover inventory has dropped 38%, freeing roughly $1.5M in working capital. Warehouse managers report spending about three hours a week on replenishment, down from twelve.
The numbers matter, but the cultural change matters more. The warehouse managers, who initially regarded "the algorithm" with appropriate suspicion, now ask for it to be extended — to procurement, to seasonal forecasting, to liquidation recommendations. The asks are how we know it's working.
Ferrum's internal team has taken full ownership. We hand off the codebase, the models, and the operations runbook in the final week of the engagement, with two months of optional Slack support afterward. The system has run for six months without us, with model retraining handled by a junior engineer on Ferrum's team.
A second engagement, scoped for next quarter, will extend the same pattern to procurement decisions — same architecture, same human-in-the-loop discipline, different decision space.