This course adopts a unified approach to building temporal intelligence, emphasizing its critical role in real-world systems that must operate reliably over time. Temporal intelligence is central to applications such as predictive maintenance, autonomous systems, finance, healthcare monitoring, and decision-making under uncertainty. The course presents a complete toolbox that integrates temporal modeling to capture system dynamics and uncertainty, forecasting to anticipate future trajectories, and anomaly detection to identify unexpected or abnormal behavior, treating these capabilities as interdependent components of a single coherent framework for building robust, time-aware intelligent systems 

Temporal intelligence systems are built from several complementary building blocks: time series modeling, causal reasoning, anomaly detection, and reinforcement learning. Time series modeling, using both classical statistical methods and modern deep learning, captures how systems evolve over time and enables prediction under uncertainty. Causal modeling extends this by representing how actions and interventions affect future states, rather than merely correlations. Anomaly detection provides self-monitoring by identifying deviations from expected temporal behavior or regime changes. Reinforcement learning closes the loop by using learned temporal and causal models to select actions that optimize long-term outcomes, transforming prediction and monitoring into adaptive, goal-directed intelligence.

World models are a central component of temporal intelligence systems, providing an internal, learned representation of how an environment evolves over time. They capture latent state dynamics and the effects of actions, allowing the system to simulate future trajectories without direct interaction with the real world. By enabling imagination, counterfactual “what-if” reasoning, and long-horizon prediction, world models unify temporal modeling, forecasting, and decision-making. This makes them especially powerful for planning, reinforcement learning, and detecting when real observations diverge from expected dynamics.

The course adopts a code-first approach, introducing temporal intelligence through hands-on implementations with a focused set of widely used libraries. Time series modeling and forecasting are covered using tools such as statsmodels, Prophet, and darts, combining classical and modern methods. Deep time series models are built with PyTorch and transformer-based architectures for long-horizon prediction. Anomaly detection is explored using PyOD and change-point detection techniques. Reinforcement learning is introduced with Gymnasium and Stable-Baselines3, while world models and model-based reinforcement learning are illustrated through frameworks such as Dreamer, demonstrating how learned dynamics support planning, imagination, and long-term decision-making.

The course syllabus is designed to enable students to begin their projects while learning the material. As the course continues, they will enrich their projects with the concepts they acquire. Each team will give several in-class presentations for discussion and feedback. 

As standard tasks are increasingly handled by AI and mature libraries, professional developers' expectations shift toward innovation and rapid integration. Accordingly, a key requirement for student projects is to tackle new use cases by generating unique data and training or fine-tuning a task-specific audio or image processing model.