No black box. Here's the data, the models, and the gate every product has to pass before it reaches you.
149M raw AIS positions (2021→now) → 335K cleaned berth calls → 19K ground-truth wait labels verified against raw trajectories. Live queue state from official port-authority feeds across 18 ports. 45M COMEX customs records (2013→now) for trade flows. Hourly weather (ERA5 back to 2021) for congestion. Independent cross-checks from IMF PortWatch.
Gradient-boosted quantile models (P10/P50/P90) for berth wait; calibrated classifiers for demurrage risk; horizon models for congestion. Always time-based train/test splits — never random for time-series. Commercial vessels only (>100m, proper cargo types).
Every product must beat a naïve baseline on held-out data or it does not ship. We killed numbered FIFO queue positions (worse than a coin flip at Santos) and multiport ML congestion where it lost to persistence. What survives is on the accuracy page, recomputed monthly.