Today, we released the NVIDIA AI Blueprint for building data flywheels—a reference workflow that automates continuous optimization of AI agents. It establishes a self-reinforcing data flywheel, using log data and up-to-date business knowledge to refine and redeploy more accurate, efficient models over time.
The data flywheel blueprint leverages NVIDIA NeMo™ microservices for model customization and evaluation.
Highlights from this release include:
- Self-improving agentic workflows—Convert production traffic into training and evaluation data and automatically trigger retraining and redeployment pipelines.
- Automated model evaluation and fine-tuning—Continuously benchmark existing and new candidate models for accuracy and ingest real-world data from prompt-completion logs to fine-tune models.
- Performance-optimized model selection—Promote smaller, faster models that match larger ones in accuracy and reduce inference latency and cost.
- Seamless deployment and routing—Admins can evaluate and seamlessly deploy and manage candidate models as NVIDIA NIM™ microservices.
- End-to-end microservice architecture—Leverage a modular, scalable system of NVIDIA NeMo and NIM microservices, purpose-built for production-grade AI agent optimization.
Get started:
📖Read our new technical blog to explore how our newest AI blueprint can help enable self-improving AI agents by automating model optimization.
📺 Watch our new how-to demo to see the data flywheel blueprint in action.
⬇️ Download the data flywheel blueprint to get started.