How a 10-engine multi-agent ML system saved Localiza $10M+ annually by autonomously pricing fleet across 950 branches
950 branches, 300,000+ vehicles, and pricing decisions being made by a spreadsheet from 2009
Localiza is LatAm's largest vehicle rental and mobility company. By the time I joined the pricing team, the fleet was massive — and the pricing logic hadn't materially changed in years. Branch managers were setting daily rates from centrally distributed spreadsheets. The spreadsheet was updated weekly. The car rental market moves hourly.
The business consequence was measurable: idle fleet time was too high in certain locations and utilization was approaching 100% in others with rates still priced at off-peak levels. We were leaving margin on the table on both ends.
The harder problem was organizational. 10,000 branch employees had been pricing cars the same way for years. Any new system had to be accurate enough to trust and simple enough to explain.
10 specialized engines, each solving one pricing signal — orchestrated into a unified recommendation
Five ML models each focus on a single pricing signal: demand forecasting, utilization prediction, competitor price movement, seasonality adjustments, and fleet repositioning cost. A separate orchestration layer combines outputs into a branch-level pricing recommendation, updated every hour.
I personally built the demand forecasting and utilization prediction models — both XGBoost, trained on 3 years of historical rental data with engineered features for local events, holidays, and competitive dynamics. The event bus processed 50M events per month at peak.