Real-time urban traffic estimation and prediction under sparse sensing, partial observability, and varying demand persist as difficult optimization issues in intelligent transportation systems. Recent centralized deep models typically require large historical datasets, incur high computational costs, and have insufficient adaptability to dynamic environments. To this end, we explored the design of a scalable and robust framework for improving traffic reconstruction and forecasting with incomplete observations. To accomplish this goal, we presented an organizational intelligence model, in which the traffic network was modeled as a dynamic graph of lightweight local computational units assigned to each road segment or intersection. These entities work together in distributed optimization processes that include adaptive influence weighting, role specialization, and structural coordination. This architecture separates local computation from global coordination, enabling efficient training without increasing model depth. We tested the framework on benchmark traffic data and realistic urban environments in Taif, Saudi Arabia. The proposed model on the benchmark dataset reduced the test root mean-squared error from 0.5578 to 0.3530 (36.7%), and the test mean absolute error (MAE) from 0.3013 to 0.1614 (46.4%) relative to a centralized baseline. The validation loss decreased by 48.0%. On real-scale deployment units, total MAE fell from 6.6473 to 6.2060, while hidden-node MAE decreased from 6.8816 to 6.4787. Other experiments demonstrated lower inference latency, increased robustness in limited-data scenarios, and stable transfer to unseen environments. These results show that organizational intelligence is an efficient, scalable, optimization-oriented alternative for next-generation real-time urban traffic management.