We will have two talks with additional time for networking.

Talk 1: Moritz Wegener (Data Scientist at ATVANTAGE): TBA

Talk 2: Gergő Szita: ICG-enhanced FOI of Rheumatoid Arthritis Classification

Follow us on LinkedIn to learn more about the talks and the agenda, and click here to secure your spot. See you in March! 🤖

Agenda

18h30 Open Doors

19h00 Welcome & Intro

19h15 Billy Kunze (AI Agent Engineer at Nandu) & Nils Zündorf (Co-Founder at Nandu): Teaching LLMs to Work with Data: From Zero to Production with LangGraph & LangSmith

LLMs are great at language — but terrible with data. Over the past year, we built a production system that bridges that gap: combining LangGraph and LangSmith with real analytics pipelines to turn raw data into contextual insights. This talk shares our hands-on learnings from starting at zero to running a full production setup in just twelve months: how we structured multi-agent workflows, managed retrieval and context, and connected the system to SQL-based sources and evaluation traces. Expect a mix of architecture lessons, pitfalls, and examples of how to make LLMs reason with data, not just talk about it.

19h50 Alexander Khachikyan (Principal AI & Data Scientist at PAYBACK) & Falko Trischler (Director Data Science & Machine Learning Engineering at PAYBACK): Intelligent Agents in Loyalty Programs: PAYBACK’S Experiences with Agentic AI

Autonomous Agents that perform tasks proactively and adapt to complex environments hold immense potential for automation and personalization. But how to bridge the gap from theory to productive reality within a complex enterprise environment? This talk offers an honest dive into PAYBACK‘s enterprise path with Agentic AI. We’ll detail what worked well, where we encountered hurdles, and the lessons learned in bringing AI agents into production across multiple domains. Through concrete use cases we’ll share first-hand insights. We’ll also demonstrate how PAYBACK leverages LLMOps and multimodal models to drive efficiency, creativity, and scalability within our loyalty programs.

20h20 Networking with food and drinks provided by PAYBACK

Join the discussion on CAIML #40 here.

We will have two talks with additional time for networking.

Talk 1: Billy Kunze (AI Agent Engineer at Nandu) & Roberto Russo (Co-Founder at Nandu): Teaching LLMs to Work with Data: From Zero to Production with LangGraph & LangSmith

Talk 2: Alexander Khachikyan (Principal AI & Data Scientist at PAYBACK) & Falko Trischler (Director Data Science & Machine Learning Engineering at PAYBACK): Intelligent Agents in Loyalty Programs: PAYBACK’S Experiences with Agentic AI

We will share more details on the talks and an agenda soon. See you in January 🤖

Click here to secure your spot.

Agenda

18h30 Open Doors

19h00 Welcome & Intro

19h15 Juri Wiens (AI Research Engineer at REWE digital) - One Voice, Many Minds: Engineering Voice-First Multi-Agent Systems

In settings where hands are busy, voice first agentic systems promise autonomous, goal oriented assistance and often need to scale as distributed multi agent architectures for extensibility, including remote agents. There is an option space, but this talk focuses on real time APIs backed by speech to speech models. We show how stateful, bidirectional audio shapes latency budgets, session state, observability, and orchestration. We outline patterns for coordinating heterogeneous agents, text and real time, in process and remote via A2A, behind a single conversational voice. Attendees leave with concise design principles, key tradeoffs, and pitfalls to avoid when scaling from one voice to many minds.

19h50 Ole Bialas (Research Software Consultant at University of Bonn) - What is “neural” about artificial neural networks? On the differences and similarities between artificial and biological intelligence

While, historically, the field of AI was strongly influenced by neuroscience and cognitive psychology, recent breakthroughs came predominantly from innovations in engineering. However, with traditional scaling laws seeing diminishing returns and models running out of new data to train on, the time seems particularly ripe to turn to neuroscience for inspiration. In my talk, I’ll present recent research on similarities and differences between human brains and artificial neural networks. For example, deep networks trained on object recognition remarkably predict neural responses across the primate visual system, with different layers corresponding to different stages of cortical processing. Yet important differences remain: while both humans and ANNs learn to categorize images, humans organize categories primarily by semantic and functional relationships, whereas ANNs rely more heavily on perceptual similarity. These comparisons inevitably raise larger philosophical questions about intelligence and agency in biological organisms and artificial systems. While I won’t be able to give conclusive answers, I hope to provide the audience with a new lens through which to view these questions.

20h20 Networking with food and drinks provided by REWE digital

Join the discussion on CAIML #39 here.

We will have two talks with additional time for networking.

Talk 1: Juri Wiens (AI Research Engineer at REWE digital): One Voice, Many Minds: Engineering Voice-First Multi-Agent Systems

Talk 2: Ole Bialas (Research Software Consultant at the University of Bonn): What is “neural” about artificial neural networks? On the differences and similarities between artificial and biological intelligence

We will share more details on the talks and an agenda soon. See you in November 🤖

Click here to secure your spot.

Agenda

18h30 Open Doors

19h00 Welcome & Intro

19h15 Tomaz Bratanic (Graph ML and GenAI research at Neo4j) - Agentic GraphRAG with MCP servers

This talk explores design patterns for integrating graph memory into agentic workflows, discusses trade-offs between retrieval accuracy and computational efficiency, and highlights how persistent knowledge unlocks more capable, personalized, and trustworthy AI systems.

19h50 Pablo Iyu Guerrero (AI Inference Engineer at Aleph Alpha) and Lukas Blübaum (AI Engineer at Aleph Alpha) - Tokenizer-free language model inference

Traditional Large Language Models rely heavily on large, predefined tokenizers (e.g., 128k+ vocabularies), introducing limitations in handling diverse character sets, rare words, and dynamic linguistic structures. This talk presents a different approach to language model inference that eliminates the need for conventional large-vocabulary tokenizers. The system operates with a core vocabulary of only 256 byte values, processing text at the most fundamental level. It employs a three-part architecture: byte-level encoder and decoder models handle character sequence processing, while a larger latent transformer operates on higher-level representations. The interface between these stages involves dynamically creating “patch embeddings”, guided by word boundaries or entropy measures. This talk will first introduce the intricacies of this byte-to-patch transformer architecture. Subsequently, we will focus on the significant engineering challenges encountered in building an efficient inference pipeline, specifically coordinating the three models, managing their CUDA graphs, and handling their respective KV caches.

20h20 Networking with food and drinks provided by lise GmbH

You can find some impressions from our event here and here.

We will have two talks with additional time for networking.

Talk 1: Tomaz Bratanic (Graph ML and GenAI research at Neo4j): Agentic GraphRAG with MCP servers

Talk 2: Pablo Iyu Guerrero (AI Inference Engineer at Aleph Alpha) and Lukas Blübaum (AI Engineer at Aleph Alpha): Tokenizer-free language model inference

We will share more details on the talks and an agenda soon. See you in September 🤖

Click here to secure your spot.

Agenda

18h30 Open Doors

19h00 Welcome & Intro

19h15 Natasha Randall (Scientific Research Assistant at TH Köln) - Generating Photorealistic Flood Predictions with Generative Adversarial Networks

The forecasting of floods is traditionally carried out using extensive, expensive numerical simulations, but the advent of AI offers new frontiers to flood forecasting, such as creating synthetic photorealistic images of floods. In this talk, we explore how effectively Generative Adversarial Networks (GANs) can be used to generate flood imagery that is not only realistic, but also provides accurate predictions of the floodwaters resulting from real-world natural disasters. We also describe methods for evaluating photorealistic images, and then use these to analyse the performance of different GAN architectures and data inputs, investigating the wheres, hows, and whys of their predictions. We thus explore AI model behaviour, taking a deeper look at what deep learning models can (and can’t!) learn about physical processes, even when they have not been explicitly encoded into the model. Overall, this talk highlights just some of the many benefits of bringing AI into the dynamic fields of hydrology and geosciences.

19h50 Mohamed Amine Jebari (Lead Data Scientist at TD Reply) - Causal convergence: the iterative blending of Algorithms and expertise

Working in the marketing analytics field often feels like an endless pursuit of the truth. While humans intuitively converge on causal relationships—linking an illness to a sick colleague, or assuming the sunrise causes the rooster to crow—establishing causality in statistics and econometrics is far more complex. To increase our confidence in causal claims, we apply a wide range of methods: randomized control trials, quasi-experimental designs, simple regressions, complex Bayesian structural models, etc. These techniques, frameworks, and tools are undeniably valuable. Yet, causality ultimately rests on human judgment. We draw on accumulated domain knowledge—built over years of experience—that enables us to form hypotheses, interpret findings, and correct course when needed. This is where experts become equal partners in the causal inference process. Domain knowledge is essential to designing interventions, defining counterfactuals, and analyzing model outputs. It helps us answer questions like:

• What exactly do we mean by “impact”?
• What mechanisms are plausible?
• Are we measuring what truly matters?

Through the iterative process between data science and human insight—testing, refining, challenging assumptions—we begin to converge towards the truth, closer to understanding the cause of phenomenons. Nevertheless, causal convergence is about the journey — a continuous back-and-forth between action and reaction, data and judgment, models and experience.

20h20 Networking with food and drinks provided by KölnBusiness

You can find some impressions from our event here and here.

We will have two talks with additional time for networking.

Talk 1: Natasha Randall (Scientific Research Assistant at TH Köln): Generating Photorealistic Flood Predictions with Generative Adversarial Networks Talk 2: Mohamed Amine Jebari (Lead Data Scientist at TD Reply): Causal convergence: the iterative blending of Algorithms and expertise

We will share more details on the talks and an agenda soon. See you in July 🤖

Click here to secure your spot.

Agenda

18h30 Open Doors

19h00 Welcome & Intro

19h15 Leonard Kunz (Data Scientist at taod Consulting) - Beyond the Hype Cycle: Constraint Optimization with OR-Tools

Using a concrete room planning problem in the context of a private university, Leonard Kunz demonstrates how complex business requirements can be formulated and solved as a mathematical optimization problem. We share our experiences with meta heuristics (including Ant Colony) in pygmo, why we ultimately chose a constraint optimization framework (OR-Tools), and which strategies eventually led us to a solution. The focus is primarily on the fundamental principles of modeling and the key lessons learned that will help us in future optimization projects.

19h50 Katharina Morik (Prof. Dr. at TU Dortmund, Speaker of SFB 876, co-founder of the Lamarr Institute) - How to introduce AI – it is all about agents

Introducing AI does not mean to be the evangelist of the newest AI technique. You better do not talk about the learning method, in the beginning. Instead, find out which action needs to be enhanced and what is the real criterium of success. Then ML becomes an agent performing the application action and you can measure the performance in terms of the real criterium, not in terms of ML loss measures. After a long time of acquiring, understanding, and massaging data, the tool and method selection is up to the ML introducer who knows their properties (correctness proofs, robustness, real-time, energy demands, privacy preserving…). Hardware selection might also be relevant. Finally, the agent is deployed, and you realize whether you have the right allies in the application field. For illustration, I’ll show how to compose an intensive care assisting agent from many classifications. As a second use case, I am prepared to show one of the use cases from steel manufacturing:

• quality assurance in interlinked manufacturing or
• managing several models for steel making (BOF) Or maybe easy listening to an insurance use case: customer churn prediction? It is up to the audience!

20h20 Networking with food and drinks provided by taod Consulting