Practical LLM Service Development for Spring Developers

Hello, I’m Lee Do-won (Kenneth).

I’ve opened the free course “Spring AI LLM Cost Monitoring — Build a Token Dashboard with One Line of Actuator”.

Why did I create this course?

The question I was asked most often while consulting on enterprise AI projects was this.

"We built a chatbot, but how much will this cost per month?"

LLM service costs vary with every request, responses can take several seconds, and the causes of failures lie outside our code. Yet many teams monitor them like ordinary web services and only discover the problem after receiving the bill.

Zero errors, Health UP, HTTP 200. An incident where input tokens increase several-fold while everything appears normal cannot be detected by conventional monitoring.

What this course does

• With one line of Actuator, expose the token, latency, and error metrics that Spring AI is already recording.

• With a single docker compose command, launch a Prometheus + Grafana token dashboard.

• We find “occasionally slow requests” using p95 instead of the average.

• With a single topK 4 → 20 setting, you will reproduce firsthand a “failure that looks normal” where input tokens increase approximately fourfold, and detect it on the dashboard.

No business code changes are needed. Just one dependency line and a few configuration lines.

Course Information

• Duration: Approximately 1 hour (including hands-on practice)

  
  

• Lab environment: JDK 21 or later, Docker Desktop, and an OpenAI API key or Ollama (free)

• Provided materials: hands-on project, docker-compose, Grafana dashboard JSON, load scripts (bash · PowerShell), and step-by-step course materials

Metric names follow the Spring AI standard, so you can use the hands-on dashboard as-is in your project. Even without an API key, you can follow along all the way using Ollama.

Recommended for those who

• Backend developers building a PoC or internal chatbot with Spring AI and preparing for launch

• Actuator and Prometheus users who don't know what to monitor in LLM services

• Team leads and architects who need to report the cost of adopting LLMs in numbers

If you want to go deeper

The custom metric that converts tokens into KRW costs, cost spike alerts, and distributed tracing (Jaeger) for viewing a single request step by step are covered in the 『Practical LLM Service Development for Spring Developers』 Section 16.

👉 Practical LLM Service Development for Spring Developers

If you get stuck while taking the course, please post your questions on the Q&A board. Questions like "I followed the steps, but the graph is empty" are welcome too.

The feedback you leave in your course review is a great help when creating the next course.

Thank you.
Best regards,
Dowon Lee

Hello. We’ve added a new supplemental lecture to the Practical LLM Service Development for Spring Developers (Part 1) course. In this approximately one-hour lecture, you’ll learn how to integrate TypeSafe AI’s recently released decision model, Jev, into a Spring Boot application.

Why was it added?
In Section 4, we covered how to use entity() to receive an LLM response as a Java object, as well as the practical pitfall of parsing breaking when the model wraps JSON in code fences. Whether you ask through a prompt or automate it with entity(), the limitation remains that you are ultimately "asking" the LLM to follow a particular format.
Jev takes a different approach. Instead of writing text, it chooses an answer from a predefined set of options and returns it along with its probability. In other words, the format is structurally guaranteed. Processing the same input in both ways makes it much clearer what to consider when designing "structured output." That is why we organized this content as an extension of Section 4.

What will you learn?

1. The Concept of Jev and System One Models: What It Means to Return a “Typed Decision” Instead of Text

2. The difference between LLMs and Jev: descriptive answers vs. multiple-choice answers, and what each does well and poorly

3. Three question types (choice · score · noul) and the /v1/systemone request structure

4. Calling from Spring: Call the Jev API with RestClient and use a Java enum to create choices and convert the response in a type-safe manner.

5. Comparison demonstration: Analyze the same movie review using entity() and Jev, then compare speed, format errors, confidence, and consistency.

Updated materials included

• Section 4 Example Code (ch4): Two Review Analysis APIs (/api/movies/review/llm, /api/movies/review/jev) and the Addition of JevClient

• Supplementary lecture slides and course materials

• Comparison test scripts: review-compare.sh for macOS and review-compare.ps1 for Windows

Things to know in advance

• Jev is not a feature of Spring AI, but rather a separate external service provided by TypeSafe AI. In the course, it is called directly using Spring AI's RestClient, not its ChatClient.

• Jev is currently in the early access stage, so there may be a wait to obtain an API key. Jev is disabled by default in the example code (jev.enabled=false), so you can proceed with the existing exercises as usual even without a key. If you don't have a key, watch the demonstration in the video and follow along with a focus on the concepts and code.

• As this service has just been launched, its pricing or API may change later. If any changes occur, we will notify you again through a new announcement.

📌 Go directly to the course: https://inf.run/VVpnV

If you have any questions while taking the newly added lessons or get stuck during the hands-on exercises, feel free to leave a comment. We’ll continue to regularly add content that you can use right away in your work.

Hello, I’m Lee Do-won.

"I’ve created Jenkins pipelines, but I’ve only heard of ArgoCD and GitOps as concepts."

"I’m at a loss as to how to design deployments for each dev, staging, and production environment."

For those facing these challenges, we have opened the “Practical GitOps CI/CD with Jenkins and ArgoCD” course on Inflearn.

Here’s what sets this course apart

• Beyond traditional Jenkins-based CI/CD, we also cover ArgoCD-based GitOps.

• We directly compare the Push (Jenkins) and Pull (ArgoCD) approaches on the same cluster. When someone makes a manual change with kubectl, you can see firsthand how one remains unchanged while the other automatically recovers, and experience the differences in how rollbacks and deployment histories are reviewed.

• Combine the two approaches to build a hybrid deployment pipeline from start to finish, with Jenkins Push for dev·staging and ArgoCD Pull (manual Sync approval) for production.

• You will practice essential deployment techniques needed for operations, including canary deployments, blue-green deployments, RBAC, approval procedures, incident response, and automatic rollbacks.

Practice Environment

We’ll proceed using just Docker Desktop. It works on both Windows and macOS, and you don’t need a cloud account such as AWS or any additional fees.

Recommended for

• Backend and infrastructure engineers who have used Jenkins but are new to ArgoCD/GitOps

• Those who want to move Kubernetes deployments from manual kubectl apply to automation

• DevOps and SRE professionals concerned with operational requirements such as deployment approvals, rollbacks, and audit trails

The total runtime is approximately 10–11 hours, and hands-on materials and a troubleshooting guide are also included.

👉 View the course: https://inf.run/dEjZE

If you have any questions, feel free to leave them in the comments or in the course Q&A!

We have newly launched the Practical LLM Service Development for Spring Developers (Part 1) course. Although it may feel a little overdue, this course covers what you need to learn to develop LLM applications using the Spring framework, which is familiar to Java developers. We hope you will take a great interest in it.

Why did I create this course?

These days, keywords like RAG, vector search, and AI agents can be found everywhere, but there aren’t many resources that address, at a practical level, “how to integrate them with a Spring Boot backend.”

This course incorporates into its curriculum the problems I encountered while debugging an actual RAG recommendation endpoint—including how chunking documents by token broke their semantic units and degraded search quality, and how I solved the problem by switching to paragraph-based chunking. Rather than simply showing you the correct code like a tutorial, we’ll follow along together to understand “why it doesn’t work → how to diagnose it → how to fix it.”

Recommended for the following people

• Backend developers with Spring Boot experience who are new to developing LLM/AI services

• Developers who want to apply RAG and vector search at the level of real-world projects, not just tutorials

• Professionals who need to directly implement internal AI features (chatbots, document search, recommendations, etc.) using Spring.

Experience creating REST APIs with Spring Boot is all you need. No prior knowledge of AI/LLMs is required.

What will you learn?

1. Spring AI Core Fundamentals Basic usage of ChatClient, structured output, and streaming response implementation

2. Building a RAG Pipeline Vector store integration, comparison of document chunking strategies and hands-on troubleshooting, search quality improvement, and prompt design

3. Tool and Multimodal Expansion Integrating external tools using MCP (Model Context Protocol), processing multimodal inputs such as images, and integrating with Dify

4. Service Stabilization and Real-World Projects Applying guardrails, designing reliable services, and a mini project that integrates everything you have learned so far

The appendix also covers how to use AI coding tools and Ollama.

Series Guide

This release is Part 1 of the full series. Parts 2 and 3 will cover more advanced architectures and real-world cases, and once the entire series is released, we plan to offer a discount on the full package.

📌 Go to the course: https://inf.run/VVpnV

Feel free to leave any questions or inquiries about the curriculum in the comments. Through Part 1, I hope to take the first step with you toward integrating an LLM into your Spring service.