Bayer - Turning Public Data Into Decisions

A single national spray recommendation is wrong somewhere in Brazil before it ships: the country grows crops across a continent, and the pest, with the weather that carried it, changes state to state. So for Bayer Crop Science I built it at the grain the decision had: one answer per area, which product and how much of it, across more than 5,550 municipalities.

Each one carried three layers: satellite imagery, reported pest outbreaks, and regional climate. The dose mattered as much as the product: too little and the pest survives the season, too much and the grower paid for chemistry the field did not need.

A second system read customs data at scale. Agrochemical shipments leave public records: what moved, how much, when it landed. I made them comparable across sources and formats, and surfaced the arrivals large enough to matter to anyone planning supply.

Both systems had the same shape underneath: the data was public and free, and worth nothing in the state it sat in. Getting it into a form a decision could stand on was the work.

I was a consultant, not an employee, so nothing I built arrived with authority of its own. It had to be right in front of people checking it against the market.

Two mathematicians from the University of São Paulo built this with me, fluent in Fortran and new to Python. The math was never the problem; the tooling was. I taught them the language while the work shipped, the only way anyone learns one they mean to keep.