Demand forecasting for automotive glass

The context

The client makes the glass that ends up in the windshields at GM, Ford and Stellantis. Planning that production depends on anticipating each automaker's demand, and that information did not arrive on its own: it lived behind a dozen client portals, each with its own personality.

The technical problem

Two problems chained together. First, half the work was getting and cleaning the data, which is the half nobody puts in the slide deck. Second, a forecast nobody opens is worth nothing, and delivery is where most models quietly die.

What I built

Autonomous bots scheduled as Databricks jobs figure out on their own how to get through the client portals and bring the data back. On top of that, time series and regression models run on Python and PySpark, with the full lifecycle covered across Databricks and Azure: data prep, training, evaluation and versioning. The results go out by themselves as daily executive reports over email, PDF and Office files.

Decisions that mattered

• →Time series and regression before anything fancier: the bottleneck was in the data, not in the model.
• →The bots figure out on their own how to get through each portal, instead of one script per portal that breaks the day the HTML changes.
• →The reports send themselves. Delivery is where most models quietly die.
• →The full lifecycle across Databricks and Azure, with versioning: it is the least glamorous part and it is why the models are still right six months later.

Results

Planning cycles got shorter and inventories got tighter. A workflow that used to eat days now runs from one trigger, and the team has happily forgotten it exists. Nobody in that office has had to ask for a report since, which I count as the real win.

My role

AI engineer and data scientist, as a consultant. I own the forecasting, the bots that fetch the data, the MLOps side, and sitting down with the business owners and with IT so the forecast is one they will actually use.

What I learned

What decides whether a model survives is not the model. It is that the data arrives without anyone chasing it, and that the result shows up where people are already looking.