The situation
A manufacturing site wanted to forecast and manage its energy behaviour at the level where decisions are actually made: individual zones and systems, not a monthly total for the whole plant.
The constraint
The data arrived from incompatible worlds. Operational telemetry, building and energy models, and years of historical consumption all described the same factory in different vocabularies, at different resolutions, with different gaps. Any model was only going to be as defensible as the pipeline underneath it.
What I decided
I built the pipeline as the deliverable and treated the models as replaceable. Heterogeneous sources were reconciled into one training and inference path, so a model could be retrained or swapped without renegotiating the data layer. I also held the interface between the engineering teams, the software developers, and the energy managers, because the forecast was worthless if the people running the plant could not say why it was telling them something.
What shipped
Production-grade machine learning pipelines feeding GPU-accelerated models and a real-time analytics layer, supporting operational diagnostics and scenario analysis at zone and system level.
What it produced
A validated pilot with a path to replication across other sites, and explainable outputs that energy managers could act on rather than merely receive.