About This Course
<div>Somebody in a meeting asks why the new AI assistant keeps inventing ticket numbers. Everyone looks at the most technical person in the room. If that person is you, and you don't have an answer, then this course is for you.</div><div><br></div><div>I built this course for the technical professionals who suddenly have AI in every product they touch: solution engineers, DevOps and platform engineers, IT pros, architects, and anyone else expected to have real answers. Why do models hallucinate? What does a token actually cost? Should your team call a hosted API or run something local? What is MCP, and why does every vendor slide deck suddenly mention it?</div><div><br></div><div>Most AI training today falls into two buckets. The beginner courses that teach you to draft emails with ChatGPT, or the engineering bootcamps assume you want to spend 30 hours writing Python. Almost nothing exists for the person in between, who needs a working understanding of the whole system more than another tool tutorial. That gap is the reason this course exists.</div><div><br></div><div>How the course works</div><div><br></div><div>We spend about five and a half hours building up one complete picture of the modern AI stack. I teach the way I always have on Udemy: short, focused lectures, with a lightboard for the concepts that need drawing out and live demos so you can watch everything actually run.</div><div><br></div><div>The whole course follows a single running example, a fictional internal support assistant called HelpDesk AI. Every section adds one layer to it. First we look inside the model it runs on. Then we price that model out and decide where it should live. We give it company knowledge with RAG, turn it into an agent, wire it to real systems over MCP, and then, in my favorite section, we attack it and watch it leak data. By the final lecture, the entire architecture is on the lightboard, and every piece of it is something you understand.</div><div><br></div><div>Along the way you'll learn:</div><div><ul><li>How large language models actually work: parameters, training vs inference, and why the same prompt gives different answers</li><li><span style="font-size: 1rem;">Tokens and context windows, and how they drive both model behavior and your bill</span></li><li><span style="font-size: 1rem;">Token economics: estimating what an AI workload really costs, and how prompt caching changes the math</span></li><li><span style="font-size: 1rem;">Choosing a model: the tradeoffs between performance, cost, and latency, plus LLM vs SLM and open vs closed weights</span></li><li><span style="font-size: 1rem;">Hosted APIs vs self-hosted models, including a demo running a local model with Ollama</span></li><li><span style="font-size: 1rem;">Prompt engineering that goes past the basics, and where context engineering takes over</span></li><li><span style="font-size: 1rem;">Retrieval-augmented generation (RAG): embeddings, vector search, and when to use RAG vs fine-tuning vs a bigger context window</span></li><li><span style="font-size: 1rem;">AI agents and agentic AI: the agent loop, tool calling, subagents, orchestrators, and human-in-the-loop controls</span></li><li><span style="font-size: 1rem;">What a harness is, and why the model is only half the product</span></li><li><span style="font-size: 1rem;">Model Context Protocol (MCP): hosts, clients, servers, and a live demo connecting one</span></li><li><span style="font-size: 1rem;">Orchestration frameworks like LangChain, LangGraph, and CrewAI, explained in plain terms so you know when you'd care</span></li><li><span style="font-size: 1rem;">AI security fundamentals: prompt injection, over-privileged agents, and why observability matters</span></li><li><span style="font-size: 1rem;">Every section ends with a quiz so you can validate these ideas yourself.</span></li></ul></div><div><span style="font-size: 1rem;">Before you enroll</span></div><div><br></div><div>You will see code in this course, because pretending AI systems involve no code would be silly. An agent created using Python. A tool call is JSON. An MCP config is a file. I put them on screen and walk through what they mean, and you will never be asked to write or debug anything. The only prerequisite is general technical literacy. If you can explain what an API is, you're technical enough for this course.</div><div><br></div><div>Fair warning on scope: we cover some of the basic security fundamentals here. The deep material on securing AI systems belongs to a separate, dedicated course. There's just too much there to include it.</div><div><br></div><div>***</div><div><br></div><div>About the Instructor</div><div><br></div><div>I'm Bryan Krausen. I've spent years teaching Terraform, Vault, and GitHub certification courses here on Udemy, and I work with this technology hands-on as a consultant. Every course I make follows the same pattern: draw the concept, demo the real thing, then hand you an exercise.</div><div><br></div><div>Words like agents, RAG, MCP, and context windows are showing up in vendor pitches and architecture reviews right now, whether you feel ready for them or not. After this course, you'll know exactly what sits behind each one, and you'll be the person in the room who can explain it.</div><div><br></div><div>Enroll and let's get started.</div>
What you'll learn:
- Explain how generative AI and LLMs work: tokens, parameters, context windows, training vs inference, and why models hallucinate.
- Estimate what an AI workload really costs, then choose the right model and where to run it: LLM vs SLM, hosted API vs self-hosted with Ollama.
- Sharpen your prompt engineering (few-shot, chain-of-thought), then add RAG: embeddings, vector search, and when to choose it over fine-tuning.
- Break down how AI agents and agentic systems work: the agent loop, tool calling, subagents, orchestrators, human-in-the-loop, and the harness.
- Understand the Model Context Protocol (MCP) and orchestration frameworks like LangChain and LangGraph, then connect an MCP server yourself.
- Spot the top AI security risks (prompt injection, over-privileged agents, weak observability) and watch a live prompt-injection attack.