Thomas Endres's Recent LinkedIn Posts

So I Completely Rebuilt a Product and Burned Through $126 Worth of Tokens in ONE DAY Following the loss of user credentials, I recreated a complete product in one day using the #ClaudeCode #AI. Follow me going through four stages - from requirements analysis, coding, the infrastructure part until finetuning in this rather non-technical article. See how we at TNG Technology Consulting can tackle such a delicate topic.

𝗛𝗮𝗽𝗽𝘆 𝗛𝗼𝗹𝗶𝗱𝗮𝘆𝘀 𝗳𝗿𝗼𝗺 𝗧𝗡𝗚 We at TNG wish you a happy holiday season and a pleasant end of the year. In addition to the predominant topic of Artificial Intelligence, which was also the focus of our Big Techday 25, we focused extensively on #robotics in 2025. Our digital Christmas cards this year were created with the help of current #AI tools and are inspired by robotics projects at TNG. In the video, we show how our TNG robots take on Christmas tasks.🎁 Here is our Christmas card: https://lnkd.in/ea_Z4i2z ✉️ You can find the background on how the digital card was created in our making-of: https://lnkd.in/ew-ZEDwz

From Idea to MVP in a few Hours with #OpenCode I am a huge fan of video games - I can waste hundreds of hours in them. So my idea was to relive my childhood and create video game posters completely with AI. This is the story of how I created a poster generator for video games in just a few hours using OpenCode, lots of #AI.

#OpenCode can control your browser with #Playwright MCP. Let's do something unconventional. Instead of writing tests or scraping data, let's order food using the power of #agentic #AI.

3 days left until Big Techday.

Today TNG Technology Consulting is at the AMD booth at the Dell Technologies Forum in Vienna speaking about #ai and #robotics. We brought our G1PO with us.

We created a showcase - a humanoid robot driving a go-kart. Now it is racing a Porsche Hypercar ...

I've recently become more cautious regarding use of #agentic #AI coders. The more I try to restrain #OpenCode, the more creative it gets finding ways to work around my restraints. The question is - will we have to use separate hardware in order to not leak private data? What do you think?

My First Skill for OpenCode (It won't be my last ...) Christoph Bergemann and I vibe-coded an #OpenCode skill to generate #AI game assets like hexagonal terrain tiles. It integrateds graphics creation into our game creation workflow for a TNG Technology Consulting workshop. This article demonstrates how easily custom skills can enhance OpenCode workflows.

Requirements Analysis Redefined with OpenClaw Once securely set up, the OpenClaw #AI #Agent system can perform agentic requirements analysis by identifying relevant documents, correlating information across them, and generating well-grounded answers to your questions. Through Slack, users can ask follow-up questions and have dynamic conversations that feel far more interactive and factual than traditional chatbots. In this short TNG Technology Consulting article, I describe how we did that.

On the Economy in a World Shaped by AI Agents: AI agents will replace platforms and software. Nobody will type a URL when their agent can find what they need across every service that exists. What will come next? In this article, I explore one possible way the economy could work in the future. #AI #FutureOfWork #AgenticAI #TrustNetworks #FutureEconomy

Teaching #AI to learn new things was never easier. We started to vibe-code a small game yesterday. During our coding session with GLM and OpenCode, we wanted to try out several Game Engines for the client - starting with MonoGame and finally settling for Pixi.js. There is however a big drawback. OpenCode doesn't know anything about Pixi.js. Whenever it wanted to create something in this game engine, it was basically lost, used WebFetch and created something that did not work. The Solution: Modern frameworks and libraries provide documentation - not only for human beings, but also for LLMs. For Pixi.js, there is https://pixijs.com/llms. There you can download 1 MB of Pixi documentation, made by LLMs for LLMs. You can download it and ask OpenCode to read it - or - even simpler, you can just ask OpenCode to download and read it: Read the PixiJS documentation from https://lnkd.in/dwydz2hK And, all of a sudden, the LLM spoke Pixi.js fluently and created an initial game from scratch. It felt like in the Matrix movie.

Just a quick reminder that while #AI #coding, working on a desktop environment and all the #agentic behaviour is very impressive, other fields are also progressing faster than ever. This is state of the art for image to 3D models now. Creating a Steampunk tank or soldier is as easy as using OpenAI image generator, then transforming this to an AI model with MeshyAI v6.

This is a build I am especially proud of. An RTX6000 within a ~10l case.

During my vacation, I visited the beautiful Meteora monasteries on top of magnificent mountain formations. I thought - if I had two hours of time, would it be possible to publish a completely #ai generated song? So I did. Here's what you need to know.

What a great day. Next one coming up in May 2026!

𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗶𝗻 𝘁𝗵𝗲 𝟭𝗘𝟵-𝗖𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 Our colleague Thomas Endres, Managing Partner at TNG, recently gave an interview to the 1E9 & Festival der Zukunft community. In his conversation with Editor-in-Chief Wolfgang Kerler, the focus was on our #AI research, particularly the development of the TNG #Chimera models and the work of our #benchmarking team. After a brief overview of the history of AI at TNG, Thomas explained the Assembly-of-Experts approach of the Chimera models. It allows us to build powerful open-weight LLMs despite limited resources, which has helped us make it into the top 10 most used models on OpenRouter, Inc.. The discussion also addressed current client requirements and areas of application for AI, as well as an outlook for the future. You can read the full interview (in German) here: https://lnkd.in/eSQyHauj

Ages ago, when I was a student, I created a web version of the board game "Catan" with some friends. Back in the days, I spent months creating this game. There were no JavaScript frontend libraries, no #AI to help me. Last week, I tried doing the same thing with the help of #Claude Code. After one hour, the complete game, including a game lobby and all the game mechanics, worked. 5 minutes later, there was a "Quintris" game. We really live in truly transformative times. It will be amazing to see how many indie games will be created in the next few months.

Creating a SingStar clone in one hour. I have long yearned for a rebirth of the Singstar franchise. Somehow, newer games were never that much fun and did not offer such a huge library as Singstar did. So today I just created my own version of this game using #Claude Code and #agentic #AI coding. You can just upload an MP3 file. It will split vocals and instruments, transcribe the lyrics with exact timings and analyze the voice pitch. You can then start a game in which you score points by singing at the same height. On top, Claude created some beautiful animations and nice visuals for the game. All of this was created in ~1 hour using #OpenCode with #Opus 4.6.

Our new Spanish branch just launched! I am looking forward to new challenges, new colleagues and new clients in Seville and beyond.

Join our live stream Innovation Hacking tour from Deutsches Museum Bonn in 10 minutes.

Realtime Action Chunking is tremendously important for Robotic Foundational Models, and we made it work!

If you are interested in joining this year's Big Techday, just let me know.