Tag: technology

Over the last weekend (14.-15. June 2025) I attended the {Tech: Karlsruhe} AI Hackathon. In general hackathons are great opportunities to learn new concepts, try out new frameworks or APIs, and get to know fellow builders. So I was looking forward to this.

The day before I thought about potential projects to work on. Since the hackathon was structured into three tracks with certain challenges, I needed an idea that was not only exciting and feasible, but also a good fit for the event.

Inspiration

Being inspired by IBM’s Project Debater from a 2019 (the pre-LLM era), I wanted to build something similar. If you have not heard about Project Debater, I can highly recommend this 60min recording.

Ideation

AI democratizes intelligence. With todays LLMs and frameworks we should be able to build a similar solution and even improve upon it in certain areas. Other than Project Debater, I wanted to simulate the entire debate solely between agents, without any human involvement. The outcome of the debate could then be reviewed and its inputs (arguments/prompts) tweaked, before the system runs the debate again. That way the system could be used to simulate different strategies and observe their effectiveness and impact on the outcome: winning or losing the debate. It could even use a reinforcement learning feedback loop to optimize itself autonomously, like AlphaZero, to discover the most effective arguments/strategies.

In real life just knowing the best arguments is not enough to win a debate. The arguments also have to be delivered well. That takes practice. To also support that aspect, the system should have an interactive training mode. Here the human can step into the debate and deliver its argument as part of the simulation.

I don’t think there is any doubt about the usefulness of it. Applications for debate technology are plenty. Debates are everywhere, especially in legal, politics, and the social sciences. I think all industries benefit form comparing multiple perspectives, contrasting arguments, and weighing pros and cons. This philosophical tool would provide a useful and valuable capability for any domain.

Implementation

At the hackathon I pitched the idea and we formed a team of five around it. As a first step we needed to find an appropriate tech stack. It seemed clear to us to have multiple agents interact with each other: one agent per side and a judge to evaluate and declare the winner. What agent framework should we use? We evaluated Agents SDK, Agno, ag2, CrewAI, LangChain, and LangGraph. Eventually we settled for LangGraph as it gave us the most control to shape the flow of information in the system. We used Lovable to generate a React frontend.

Our project competed in the “AI Applications” track of the hackathon. As a requirement we had to use at least three partner/sponsor technologies. That was not difficult for us. We used Mistral AI to generate arguments and evaluations. We used ElevenLabs to generate audio representations of the responses. For the interactive training mode we used Beyond Presence to give the opponent machine a human face. This created an immersive debate experience, almost as real as an online meeting.

Presentation

After two days of intensive work, it was time to presented our project. The first presentation was in front of a panel of partners/sponsors. In our track we made it to the top three finalists. This got us a second chance to present it to the entire audience of all attendees. Before the second presentation we found ourself in a situation where our application did not work anymore. Without our demo we would not stand a chance in the finals. Luckily we managed to get it working again just in time for our big presentation. And we actually won the “AI Applications” track!

Thanks to my team for bringing this idea to life. Thanks to {Tech: Europe} for organizing the hackathon.

I learned a lot and enjoyed the weekend. There is no debate about it.