In our AI Podcast: Medienwerft AI Solution Engineer André Osyguß

Let’s make it AI – but how? Everyone agrees on one thing: artificial intelligence is the future – especially for companies in e-commerce. Whether in the existing technology stack, in a current SAP environment, or in the orchestration of complex systems – the question remains: how can AI be used in such a way that processes are not only simplified but also automated?

#ARTIFICIAL INTELLIGENCE #MARKET & INNOVATION ‐ 15. September 2025

Worth listening to: In a three-part podcast series, André Osyguß, AI Solution Engineer at Medienwerft, highlights modern approaches – from a proprietary embedding model to autonomously operating AI agents to the question of the explainability of AI decisions. Language models such as ChatGPT are often perceived as a “black box” – their answers seem plausible, but how they originate remains difficult to understand.
This is precisely where companies with their own AI applications are increasingly reaching their limits: the traceability of individual decisions is not only crucial for trust, but also necessary for optimizing models in a targeted manner, reviewing results retrospectively, and consciously controlling processes as needed.

Find lots of insights and tips in the three short podcast episodes. And if you would like to learn more about the potential of AI in your company, feel free to contact us!

Part 1: More than just math The FISMWASP Multi embedding model

Learn all about this in the first episode:

What is an embedding model?
What can an embedding model do?
Why does it make sense to use your own embedding model?

Part 2: AI agents When chatbots are no longer sufficient

Find out in the second episode:

What comes after the AI chatbot?
What are the limitations of AI agents, and where is the technology currently headed?
Will AI agents be the future of automation in companies?

Part 3: Explainability AI and the issue of trust

Find out in the third episode:

What does explainability mean in the context of AI?
What’s behind the “black box” of AI – where do ChatGPT’s answers actually come from?
What are future use cases for explainability?

With an embedding model like our FISMWASP, we can act very, very quickly because it is also incredibly efficient.

André Osyguß
AI Solution Engineer at Medienwerft

AI in action:

Image search for STARK Germany

Challenge:
The range of tiles offered by STARK Germany is so extensive that conventional filters such as color or shape are often insufficient to find a specific tile. Especially when customers already have a specific idea in mind, it is easier to take a photo of a tile and search for it directly in the shop. However, tiles are a difficult case – many look very similar, and light conditions and many other factors also play an important role when it comes to finding the same tile as in the photo.

The solution:
With an AI-based embedding model, data is mapped into vectors. Each image is given a numerical representation that describes its essential characteristics such as shapes, colors, or structures. These vectors can be compared in a multidimensional space: if two image vectors are close to each other, this means that their content is very similar; the further apart they are, the more dissimilar the images are.

This creates a “map” for the embedding model, enabling AI-supported image searches: Instead of searching only by file name or keywords, the model recognizes similarities in content between images and thus delivers more precise and relevant results—even when it comes to white tiles!

To the success story