Why does a large language model write one token at a time, and how can a small drafter make it faster without changing the answer? Explore a real recorded run on an NVIDIA L40S through bets, interactive blocks and a head-to-head race. No GPU reservation needed: start anytime, finish in about 30 minutes.

Give your AI assistant direct access to your source of truth. In this hands-on lab, you'll connect an AI client to a live NetBox instance using the NetBox MCP server, then query devices, IP addresses, and sites in plain English. Learn how MCP tools turn natural-language questions into accurate, real-time infrastructure answers.

Run Hermes Agent as an on-call operations team for a storefront outage. Fan out parallel investigators, adjudicate a risky change with a Mixture-of-Agents board, execute the fix under a goal contract with machine-verified proof, and review the agent run afterward.

In this lab you build two agent architectures from scratch in Python, with no framework hiding the mechanics: an orchestrator that routes work to specialized workers, and a ReAct loop that reasons, acts, and observes until it has an answer. You'll expose the strengths and weaknesses of these architectures first hand. You'll leave understanding how each pattern works under the hood, and gain a feel for how and where these patterns are best applied.

The learner is guided through the installation, configuration, and operations of the Ansible Dev Tools MCP Server. The lab guide walks through using the Ansible Dev Tools MCP server to build an ansible role, validate it against ansible-lint rules and best practices, and ultimately publish it as trusted content.

This hands-on lab provides a comprehensive introduction to NVIDIA Run:ai, a powerful platform for managing AI workloads on Kubernetes. Designed for AI practitioners, data scientists, and researchers, this lab will guide you through the core concepts and practical applications of Run:ai's workload management system.

NVIDIA's AI Blueprint for RAG is a foundational guide for developers to build powerful data extraction and retrieval pipelines. It leverages NVIDIA NeMo Retriever models to create scalable and customizable RAG (Retrieval-Augmented Generation) applications. This blueprint allows you to connect large language models (LLMs) to a wide range of enterprise data, including text, tables, charts, and infographics within millions of PDFs. The result is context-aware responses that can unlock valuable insights. By using this blueprint, you can achieve 15x faster multimodal PDF data extraction and reduce incorrect answers by 50%. This boost in performance and accuracy helps enterprises drive productivity and get actionable insights from their data.

Routing LLM Requests with NVIDIA Switchyard teaches config-driven LLM routing through a hands-on Docker Compose lab. Students deploy an OpenAI-compatible proxy plus dashboard, use passthrough routes, binary and category-based LLM classifiers, then measure real cost, latency, and quality tradeoffs between routed and frontier-baseline models across six progressive modules.

Watch a multi-agent AI system run network operations on a live Cisco leaf-spine fabric. It detects configuration drift against a golden baseline, diagnoses root cause from live telemetry, and proposes Ansible remediation — but executes nothing without your approval. Work five realistic incidents, from a fat-finger mistake to a deliberate attack, and judge where agentic automation belongs.

A hands-on lab where learners configure LiteLLM as a local reverse proxy that intelligently routes requests between multiple LLM backends — a general model for simple prompts, a coding model and a reasoning model for complex ones — instead of hard-coding a single model into an application.