How do you keep your AI assistant on task?

How do you enforce safety and compliance policies on generative AI applications? Learn how to deploy NVIDIA NeMo Guardrails on Red Hat AI to inspect model inputs and outputs in real time. Managing generative AI risks requires controls that sit outside the underlying model. This video demonstrates how NVIDIA NeMo Guardrails integrates into Red Hat AI through the TrustyAI Operator to evaluate prompts and responses before they reach users. Using a pizza-ordering assistant scenario, see how input detectors catch prompt injections and restricted requests, while output detectors filter sensitive data like Social Security numbers and credit card numbers. Learn how administrators use Kubernetes ConfigMaps to define declarative policies for vLLM endpoints, maintaining policy enforcement regardless of model changes. 00:00 Understanding generative AI safety and operational risks 00:42 How input and output detectors inspect requests 01:18 Deploying NeMo Guardrails with the TrustyAI Operator 01:52 Pizza agent demo: Testing model safety controls 02:35 Backend configuration and Kubernetes ConfigMap setup 03:10 Configuring regular expression and sensitive data detection 03:50 Defining input and output execution flows 04:20 Quickstart resources and next steps Explore how to implement AI safety controls on Red Hat AI: 🛡️ Try the AI guardrails quickstart → https://red.ht/F5AIQS 🍋 Try the lemonade stand assistant quickstart → https://red.ht/lemonade-quickstart #AIGuardrails #RedHatAI #NVIDIA #TrustyAI #OpenShiftAI #AISafety