Where the hype ends and the ROI begins.

The operations manager was running replenishment on weekly spreadsheet reviews and gut instinct. Stockouts on the top-20 revenue SKUs were running at 14% during peak demand weeks. He knew the number was bad. What he didn't have was a forward-looking signal, a way of knowing on Tuesday that Thursday's demand would spike, before the shelves emptied. The estimated cost of stockouts was ₹78-82 Lakh per quarter, but that number was built on assumptions the team had made up because nobody had a better method.

A demand-sensing model built on 22 months of store-level daily sales data, with the previous two months excluded because they overlapped with an atypical promotions period that would have distorted the training data. The model used XGBoost, integrated with the retailer's Tally-based ERP via a CSV export bridge because the ERP's API limitations made a direct integration impractical. The output was a weekly replenishment recommendation report, not a black-box instruction. Store managers were trained to override with a documented reason, which created a feedback loop that improved the model over time.

This reflects the type of challenge our consultants are built to solve, drawn from real industry experience. Three CNC machining lines were averaging 14.2 hours of unplanned downtime per month per line. The plant manager had already increased PM frequency. It hadn't helped. Two of the three Q2 breakdowns had occurred within a week of a scheduled maintenance check, which was the number that finally made him question whether the problem was prediction rather than frequency. At ₹1.6 Lakh per hour of downtime, the three lines were costing roughly ₹68 Lakh per year in unplanned stops.

Vibration and temperature sensors were deployed on 11 critical wear points across the three lines. Eighteen months of maintenance logs were manually digitised and mapped to sensor readings. A gradient boosting model was trained on 15 months of data, validated on the remaining 3, and deployed to generate 48-72 hour advance alerts. The false positive rate was 22%, one in five alerts was a non-event. That's acceptable when the cost of ignoring a real alert is ₹1.6 Lakh. Maintenance staff were trained to treat amber alerts as a trigger for a targeted inspection rather than a full shutdown.