IBM and Confluent Release Real-Time Time-Series AI Model

IBM and Confluent have launched a time-series foundation model (TSFM) for real-time streaming data in Early Access on Confluent Cloud. This integration performs inference directly where the data moves, eliminating the latency and costs associated with loading data into separate storage and batch processing.

Real-Time Inference and State Management

This solution leverages Apache Flink to support stateful inference on streaming data. Since Flink manages the history for each time-series sequence, predictions, anomaly detection, similarity search, and optimization can be performed in real time without separate database queries. This is particularly useful for business cases where the value of a signal drops sharply over time, such as pump failure or fraud detection.

Accessibility and Scope

The IBM Granite Time Series model generalizes to unseen time-series data through pre-training. Even without being a data science expert, demand planners, fraud analysts, and process engineers can directly invoke and use the model's capabilities. It is currently accessible on Confluent Cloud on AWS, with support for on-premises and hybrid environments via Confluent Platform planned for the future.