Analytics & Machine Learning

A healthcare facility had given one supplier preferred status and suspected it was not getting the pricing that came with it. We reconstructed months of purchasing from raw transaction records and benchmarked the supplier against its own cheapest price rather than a market rate, so the findings could not be dismissed as market movement. They went into the renegotiation with a line-by-line list of every item they had overpaid for.

The system recording what came in had never been connected to the one recording what went out, and the two used different names for the same items. Joining them credibly took automated matching followed by manual reconciliation: the difference between a real figure and an invented one. Management got a deliberately conservative loss figure, built to survive being challenged.

Ordering was being done on instinct and recent memory, so fast-moving lines ran dry while slow ones tied up cash. We built demand forecasting off their own sales history and turned it into something usable at the counter: what to reorder, when, and which regular customers are due back.

Standard promotion reporting counts revenue that would have arrived anyway, which is how a campaign looks successful while quietly destroying margin. Separating the two means comparing against customers who never got the offer, not against the same customers' quieter months, since the people who redeem offers are the heaviest spenders to begin with. The result is a dashboard a marketer can drive without writing code.