Model Validation Across Predictive and Generative AI
- Observability
- H2O Driverless AI
- Enterprise h2oGPTe
- H2O Eval Studio
This lesson covers model validation across both predictive and generative AI. It starts in H2O Driverless AI, showing expert settings like the compliant pipeline and monotonic GBM recipes built for regulated industries, along with the automatic, customizable documentation generated for each experiment. It then moves to Enterprise h2oGPTe, covering collection level guardrails, PII redaction, and prompt template management. It closes in H2O Eval Studio, showing an inline evaluation of a single RAG response for fairness and bias, and a full suite comparing which models pass or fail a use case specific test like a bank teller exam.
Key takeaways
- How Driverless AI's compliant pipeline and monotonic GBM recipes build interpretability and monotonicity into a model for regulated industries
- What automatic experiment documentation in Driverless AI provides, and how regulated customers customize it for model risk management
- How Enterprise h2oGPTe applies guardrails and PII redaction, using Presidio, PII labels, or a custom regex, before content reaches storage
- Why an inline evaluation on a single RAG response shows fairness and bias scoring alongside retrieval traceability and reproducible code
- How H2O Eval Studio's evaluation suite compares models against a use case specific test, showing which prompts and models pass or fail
Start owning your AI