Willem Meints

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

Want to know when tools are your secret weapon versus when they'll sabotage your application?

Read Chapter 6 to discover the practical strategies for building LLM tools that actually work. Get the book here: leanpub.com/effective-ll...

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

When you ask an application like ChatGPT to search the web or generate an image, you're not seeing LLM magic. You're witnessing tool calling in action. The LLM detects what tool to use, extracts the necessary data, and gives your application the information it needs to make the tool call for the LLM

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

🔧 What if your agent could actually DO things instead of just talk about them?

LLMs can't access the internet, generate images, or interact with your business data—but they can call tools that do all of these things through Semantic Kernel functions.

Kids and wife are on the beach. So, its time for a book and a route check. And a coca cola zero.

Keychron K3 QMK Wireless Mechanical Keyboard (Version 3) The Keychron K3 Version 3 is an ultra-slim QMK wireless mechanical keyboard with an 75% layout. It supports Bluetooth 5.2, wired connectivity, and key remapping or macros via the Keychron Launcher web...

Funny story: Now my wife wants a programmable keyboard too. Just as long as I program the additional functions into it. And she wants a slim, and low profile format. She really liked this one: www.keychron.com/products/key...

So.... Am I useless yet? did we achieve AGI today? did we? did we?

I just figured out that there's another option to change the lighting of the keyboard based on what layer you're on. It includes a splash effect.

Yeah, I still have a roll of the stuff lying around the house. It's great for binding cables that you need to move later! :D

Customize your Keychron Q3 Keyboard QMK Firmware Full admission here, I am a keyboard nerd. I love tinkering with custom keyboards. So it's time to modify my trusty Keychron Q3 keyboard with new firmware.

I'm having a bit too much fun during the first day of my vacation. Customizing one of my keyboards: fizzylogic.nl/2025/08/08/c...

And thanks to tools like Claude I have a much easier time figuring out how to modify my keyboard. I will write a short blog post explaining the things I did to modify this keyboard.

This involves building my own firmware, which sounds a little scary, but isn't bad at all. There are excellent samples available online.

Achievement unlocked. I can change the colors of individual keys on my Keychron Q3. I totally need to get a keycap set for it that is more transparent so I can see the colors more clearly.

Vacation! Time to hack something weird. Such as the firmware of my Keychron Q3 keyboard. Apparently you can individually change the color of the leds. I have to try that in QMK.

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

Read Chapter 5 to discover the complete testing and monitoring approach—from deterministic pattern checks to model-based evaluation techniques, plus the OpenTelemetry setup that captures real production data for continuous improvement.

🚀 Grab the book at: leanpub.com/effective-ll...

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

Here's another thing I learned building LLM-based applications: Monitoring is essential because even with great tests, there's no place like production when it comes to agents and AI automated workflows.

Want to know the testing strategies that actually work for non-deterministic AI systems?

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

But here's the warning nobody talks about: when you use an LLM to test an LLM, you're not getting real scores—you're getting tokens in a sequence that looks like evaluation metrics. It's weird, but research shows these "fake" scores have remarkable agreement with human experts.

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

But here's what most developers miss — you CAN test for patterns and properties that should always hold true. Property-based testing combined with model-based validation. Instead of checking if the LLM outputs exactly "Recipe: Pasta with tomatoes," you test whether the response contains patterns.

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

🧪 Think you can test LLM applications the same way you test regular code?

Here's the brutal truth: LLMs give you different responses every time you call them. You can't test for specific outputs like traditional software.

My agenda was empty at the start of the week. Now it's completely booked. Sorry, peeps, I can't have anymore meetings. Trying to finish up before the vacation is here.

Yeah, that's at least a little bit better. Microsoft did a great job in that regard with nuget. There they have build.directory.props and build.directory.targets to get a bill of materials, but it's standardized, so you know what to look for as a tool builder.

Using the Quarkus Kubernetes extension to deploy the podcast app to a proper Kubernetes environment with all the niceties that it offers. Only to find out that I need to be more careful with config properties ;-)

Too bad, I was hoping that there would be a 99% correct pattern. Ah well, It was a nice idea. Maybe they'll come up with a better solution in the future. It sucks having to go to websites and copy/paste stuff.

@mthmulders.bsky.social My idea is this: When I ask to add a dependency, the tool checks if there's a <group>-bom/dependencies/... or a <artifactId>-bom/dependencies/... available and notifies the user that there's a BOM and asks for confirmation to use the BOM if it isn't already in the maven POM.

I can be quite complicated, but is there a pattern to the madness? I'm thinking: Is it possible to detect when you should be using a BOM?

I prefer to use claude code these days because I can run it from the terminal with cat application.log | claude -p "Check for any errors related to this one thing and fix them". In short, I can pipe in content from a file and then ask claude to process it.

Creating a small tool to help me quickly find and add dependencies to my maven projects. Or rather, claude code is working on it. I am watching a video to learn more about configuring my dygma defy a bit more.

Building effective LLM-based applications with Semantic Kernel Learn to build production-ready applications with Large Language Models using C# and Semantic Kernel. A practical guide with proven patterns and real examples.

I just updated the dependencies for the current Python samples to the latest version. Enjoy!

And go get that book: leanpub.com/effective-ll...

I will have to work on a solution for testing the Python samples in my book "Building Effective LLM-Based Applications with Semantic Kernel". You can't really find syntax problems until you run the application, so I guess I'm going to have to include unit tests with the samples.

I'm curious if they took on your offer to have a discussion about the topic. I think it's a great opportunity to explain how it works exactly and what it does and doesn't solve for them specifically.

There is nothing bad about the product in that case, but it says a lot about how mature an organization is in adopting extensive technology like Aspire.