-black){kind=icon}
A free, hands-on series that takes you from "what is AI, really?" to a working AI agent in Java, one episode at a time.
Every episode has slides, and from Episode 3 on, every episode has a Java lab you can run yourself. You don't need any AI background, a GPU or a paid API key: the labs run on a free local model through Ollama.
You'll learn: how LLMs work · calling an LLM from Java · system prompts and structured JSON output · tool calling · MCP (Model Context Protocol) with the Java MCP SDK · RAG with embeddings and a vector store · building a real agent with LangChain4j and Spring Boot.
Episodes
| # | Episode | Type | Slides | Code / Lab |
|---|---|---|---|---|
| 1 | The AI Map — AI vs ML vs neural networks vs GenAI vs LLMs | Foundations | Open slides | Teachable Machine lab |
| 2 | How an LLM actually "thinks" — tokens, weights, embeddings, next-token prediction | Foundations | PDF · PPTX | Tiny LLM Lab (browser) |
| 3 | Talking to an LLM with code — your first API call in Java, with Ollama | Java lab | Open slides | ep03 code |
| 4 | Giving the LLM hands — tool calling | Java lab | coming soon | ep04 |
| 5 | MCP: plug-and-play tools — the official Java MCP SDK | Java lab | coming soon | ep05 |
| 6 | RAG: giving the agent your knowledge — embeddings and vector search | Java lab | coming soon | ep06 |
| 7 | From demo to real agent — LangChain4j + Spring Boot | Java lab | coming soon | ep07 |
Each episode folder has its own README with what it covers, what you need, and how to run it.
How the series fits together
Episodes 1 and 2 explain what a model is: weights learned from examples, then frozen, predicting one token at a time.
Episode 3 shows that calling an AI is just sending JSON to a URL, and that the model is stateless: the model doesn't remember, your app does. Everything the model knows during a request has to be inside its context window.
From there, each episode puts something new into that window:
Context window · everything the model sees
├── System prompt ............ Ep 3 the standing rules
├── Live data from tools ..... Ep 4 get_order(4521) returns the real order
├── Tools from anywhere ...... Ep 5 one standard (MCP) instead of custom glue
├── Your knowledge ........... Ep 6 only the paragraph that matches (RAG)
└── All of it, production-ready Ep 7 LangChain4j + a web chat UI
Tools used
| Purpose | Tool |
|---|---|
| Language | Java 17 or newer (records, text blocks and the built-in HttpClient) |
| Build tool | Maven |
| IDE | IntelliJ IDEA Community (free) or VS Code with the Java extensions |
| LLM API | The official Java SDK of your chosen provider (Anthropic, OpenAI and Google all have one) |
| Free/local option | Ollama, over plain HTTP or through Spring AI / LangChain4j |
| Tokenizer demo | An online tokenizer playground; GPT-Tokenizer for token counting in Java |
| MCP | Official MCP Java SDK, Claude Desktop as the host, MCP Inspector for testing |
| RAG | An embeddings API, an in-memory store first, then pgvector or Chroma |
| Framework (Ep 7) | LangChain4j |
| UI (Ep 7) | A Spring Boot web endpoint with a simple HTML chat page |
Quick start
git clone https://github.com/kagitha-dev/zero-to-ai-agent.git
cd zero-to-ai-agent
# 1. Install Ollama from https://ollama.com/download, then get the model used in the labs
ollama pull qwen3.5:4b
# 2. Build every Java lab
mvn compile
# 3. Run one (from inside its folder)
cd episodes/ep03-talking-to-llm-java-ollama
mvn -q exec:java -Dexec.mainClass=airadar.Step1_FirstCallOr open the repo in IntelliJ IDEA (File → Open → pom.xml) and press Run next to any main method.
Repository layout
zero-to-ai-agent/
├── pom.xml Maven parent: one module per Java episode
├── episodes/
│ ├── ep01-the-ai-map/ slides (HTML)
│ ├── ep02-how-an-llm-thinks/ slides (PPTX + PDF); lab lives in its own repo, tiny-llm-lab
│ ├── ep03-talking-to-llm-java-ollama/ slides (HTML) + Java lab
│ ├── ep04-tool-calling-java/
│ ├── ep05-mcp-java-sdk/
│ ├── ep06-rag-java/
│ └── ep07-langchain4j-agent/
└── LICENSE
Follow along
- ⭐ Star this repo to get notified when the next episode's code lands.
- Questions or a bug in a lab? Open an issue.
- More from the author: kagitha.com
Author
Prasad Kagitha · @kagitha-dev
Licensed under the MIT License. Slides and labs are free to use for learning and teaching.