The airport already ran a well-tuned allocation engine for stands, gates, baggage belts and check-in counters, driven by business rules and weighted KPIs. Plans still changed through the day, aircraft were towed between stands more often than planners wanted, and many of the preferences experienced staff applied had never been written into any rule.

The airport wanted to know whether a learned model could improve on that engine without replacing it, measured on its own KPI definitions.

We trained six SANS.AI models on about 18 months of the airport's allocation history, more than 5 million operational records. They cover stand and gate prediction, passenger demand forecasting, belt and counter allocation, and an optimizer that balances every KPI at once.

The models learned the airport's working rules from its own decisions, including which carriers favor which stands and how wide-body traffic is routed, without anyone writing them down. Hard constraints such as wingspan limits and maintenance closures still override anything the models learn.

For every flight in a holdout period, we compared the SANS.AI recommendation with the allocation the existing system actually produced, scored on the airport's own KPI definitions. Planners see each recommendation with a confidence score and the reasons behind it, and nothing takes effect without their approval.