If you’re working with streaming data, you can set up a data flywheel to continuously improve your AI agents - no need to wait for large offline batches.
Yes - flywheel jobs can be configured to consume streaming events, buffer into mini-batches, and trigger retraining jobs periodically/ per business need. These flywheel jobs may include fine-tuning, in-context learning, and evaluation experiments.
You may trigger flywheel jobs (refer to the data flywheel blueprint) periodically, driven by events like:
- Model Updates: Launch flywheel jobs when a new model or NIM microservice is available, ensuring your agents always use the latest version.
- Resource Optimization: Schedule jobs during periods of low utilization to maximize compute efficiency.
- Continuous Learning: Maintain a loop where new streaming data is regularly incorporated, keeping models fresh and aligned with real-world usage patterns.
This can be useful for systems that receive a continuous stream of data such as edge devices, real-time trading data, continuous ticket triaging and classification and more.