Intelligent transportation systems (ITS) for public transit have largely been built around schedule planning, with delay and crowding prediction treated as secondary. This paper introduces the AI Public Transport Predictor, a multimodal spatio-temporal framework that jointly forecasts bus arrival delay and the number of waiting and on-board passengers. The model fuses static schedule data (GTFS), real-time vehicle feeds (GTFS-Realtime/AVL), automatic passenger counts, weather, and event information, routing each modality through a modality-specific feature pipeline into a shared fusion layer, and then into dual prediction heads for delay and crowding linked by a learned coupling term. Concept drift is handled online by a Page-Hinkley change-point monitor. On a six month GTFS/GTFS-Realtime case study, the coupled model attains a delay MAE of 1.8 ± 0.2 min and a crowd MAE of 12.3 ± 1.1 passengers, a 15–20% improvement over reproduced ArrivalNet and DeepTTE baselines and over single-task variants, with a CRPS of 2.6 against 3.4 for an LSTM baseline. The online-learning variant limits the error spike caused by an abrupt demand shift from roughly one month to approximately two hours. We also describe a deployment architecture that sustains about 2000 cached predictions per second at 50 ms latency, together with its security, privacy, and validation considerations
 

Index Terms—bus delay prediction, passenger crowding, graph neural networks, spatio-temporal forecasting, multimodal data fusion, explainable AI, intelligent transportation systems.