Using Machine Learning to Cut Logistics Costs by ₹13 Crore a Month

When an order ships in three boxes from three warehouses instead of one box from one, the customer waits longer and the company pays more. Multiply that across a 1,000Cr+ GMV D2C operation and "shipment splits" stops being an operational footnote and becomes a multi-crore monthly leak.

Framing it as a prediction problem

The instinct is to solve placement with rules: keep fast movers everywhere, slow movers central. Rules are brittle. Demand shifts by geography, season, and promotion, and static rules cannot keep up.

We reframed inventory placement as a machine-learning problem. The model learned where demand for each SKU was likely to originate and positioned stock so that the maximum share of orders could be fulfilled from a single node. Crucially, we wrapped the model in guardrails — capacity limits, safety stock, and cost constraints — so its recommendations were always executable.

What changed

• 75% reduction in shipment splits — far more orders fulfilled from a single warehouse.
• ₹13 Crore per month in logistics savings.
• Zero additional headcount — the system replaced manual judgement, it did not add to it.

Why it worked

Two things. First, the model was trained on the messy reality of demand rather than a tidy set of rules. Second — and this matters more than people admit — the ML output was constrained by operational guardrails, so the warehouse teams trusted it and acted on it.

AI projects in operations fail when the model is technically impressive but operationally ignored. The savings here came from making machine learning boringly executable.