Demand sensing across 42 manufacturing plants
$14M annual inventory carry reduction
A global industrial manufacturer had rich plant telemetry and no shared decision layer. Mercid built the lakehouse foundation and demand-sensing models that turned 42 plants' signals into inventory action.
$14M
Annual inventory carry reduction
< 1s
Telemetry latency from plant to lakehouse
98.6%
Data quality contract pass rate
42
Plants on one decision layer
Forty-two plants ran on heterogenous MES and ERP systems, each with its own data dialect. Demand planning relied on monthly spreadsheets, so inventory was carried against worst-case buffers. The CDO had funded a data platform initiative that had consumed two years and produced a warehouse nobody queried. The board had lost confidence that a single trusted number was possible.
We started not with the platform but with the decision: where would one trusted, fast number change behavior? We selected three high-value use cases — demand sensing, supplier risk scoring and plant OEE benchmarking — and designed the lakehouse specifically to serve them. A medallion architecture on Databricks, with quality contracts enforced in CI, meant every downstream model could trust its inputs without a human reconciliation step.
Streaming pipelines ingested plant telemetry in under a second; a semantic layer defined the metrics once and served them to every consumer. Demand-sensing models, retrained nightly against two years of history, replaced the monthly spreadsheet buffer logic with dynamic safety stock. The board now reviews the same inventory, demand and risk numbers the plants act on — produced by one governed pipeline.
Timeline
Three quarters to first model in production; six to steady state.
Engineering Stack
Facing a similar constraint?
Tell us the system that is slowing you down. We will return an architecture and a delivery plan.