Agenda Machine Learning Week Europe 2025

17 - 18 November 2025, Berlin

AI is evolving from answering questions to solving problems autonomously. This keynote reveals how to build AI agents that plan, decide, and act, transforming industries from software development to customer service. Learn practical frameworks for creating trustworthy autonomous systems, proven strategies for user adoption, and a roadmap from simple assistants to fully autonomous agents. Discover why the organizations mastering AI agents today will define the next decade of innovation.

Manual video review for compliance is slow and costly. To solve this, RTL built merm:ai:d – a modular AI platform for automated video understanding, classification, and enrichment. It has reduced moderation times by up to 80% and unlocked new automation potential. This session shows how RTL designed scalable AI workflows that evolved from a single use case into a strategic asset – and what it takes to align AI innovation with business value in media contexts.

The presentation will focus on the justification, development and implementation of an advanced control strategy for automating the dilution process in fill-finish manufacturing. The novel strategy includes the implementation of sensors, real-time data processing and modeling, combined with immediate process feedback control. This approach was implemented at Johnson & Johnson in the commercial facilities and allows to reach the target protein concentration with high accuracy despite process irregularities.

Understanding data is hard. Visualization techniques help us to see the meaning in data and convey it to out audience. Neural networks can be tools to extract meaning from data and turn data into beautiful representations of its underlying structures. The session will be a hands-on demonstration on how to use visualization tools from clustering over self-organizing maps to deep learning-based variational autoencoders with the aim to create meaningful and beautiful data visualizations.

Nils, Senior Data Engineer, and Justin, Senior Data Scientist, present a Case Study from publisher Axel Springer. They explain how isolated predictive use cases evolved into a next-gen marketing automation platform that speeds campaign delivery and drives subscriber growth. Justin covers their journey of ML use cases in marketing; Nils details data integrations that formed a unified Customer Data Chip and seamless automation platform. This session showcases a blueprint for marketing enablement.

Tight and precise delivery promises are fundamental at Zalando. In some countries, however, our ML models led to unwished overly broad promise windows. Postal codes historically exhibit different delivery behaviors, justifying a move from a country to a regional-level. However, clustering methods are not available for partially observable time series data structures. By extracting empirical probability distributions to cluster postal codes, we obtained regions having tighter promise windows.

Generative AI has captured the imagination of industries worldwide, widely hailed as a catalyst for innovation and efficiency, yet adoption remains limited. 

In this session, I will share how Appriss Retail combines AI approaches to address fraud and abuse in retail sales and returns. Generative AI assists analysts in exploratory data analysis and investigating organized retail crime. Meanwhile, traditional machine learning models drive real-time decisions to mitigate returns fraud and abuse

In this case study, we present the collaboration between Merantix Momentum and Cornelsen to develop AI Material Designer https://www.cornelsen.de/digital/ai, a tool that uses LLMs to generate curriculum-aligned teaching content. The talk will focus on the challenge of evaluating AI-generated materials, compare manual review with LLM-as-judge approaches, and share key lessons and business insights from deploying the evaluation strategy in an educational setting.

This session reveals how Alumil and Smiling Machines transformed a traditional industry through AI—not by starting with models, but by building the foundation first. Over three years, we reshaped culture, redefined processes, secured networks, and cleaned data flows before applying machine learning. Participants will learn that behind every AI success lies a deep transformation in mindset, systems, and continuous improvement.

Machine unlearning – the process of removing specific data from trained models – is becoming critical to comply with regulations like GDPR and the EU AI Act, and to reduce harmful biases or errors caused by flawed data. While research has advanced, deploying unlearning techniques in production remains highly challenging. This session explores practical approaches, common pitfalls, and hands-on solutions using the Vision-Unlearning Python library.

Together we’ll explore how Agentic AI is changing the way software companies create, deploy and deliver solutions. We’ll dive into business models introduced by Agentic AI, net new, emerging, but also iterative – meeting software companies half way. We’ll look what’s beyond the horizon and dive into the topic of Autonomous AI, and models enabled by this shift such as Autonomous Business Capabilities.

In this visionary round table discussion, we will dive into the emergent field of Quantum Machine Learning. Venturing beyond mere prediction, we’ll explore the transformative potential of quantum computing to speed up processing, increase efficiency, and solve complex problems that are currently beyond reach. Let’s collaboratively envision a future where Quantum Machine Learning revolutionizes industries, from healthcare to IT. The conversation is far from over – it’s just beginning.

We will explore the following questions:

1. What are the prerequisites for understanding the costs associated with the generation, storage, processing, maintenance, and sharing of data assets?
2. Which approaches are effective and feasible in determining the value contribution of data assets?
3. How can data asset valuation be integrated into organizational decision-making?
4. In what ways can a solid understanding of data-related cost drivers support proactive and informed cost control

GenAI is reshaping retail among many other industries. This session will offer an invaluable look into Amazon’s real-world GenAI integration, moving beyond theory to share the practical triumphs and hurdles. Participants will discover tangible strategies for deploying sophisticated AI at scale and learn how to spark significant product and organizational evolution, taking away key insights on navigating integration complexities to unlock transformative customer experiences and business growth.

This session explores how at PTTEP (a petroleum exploration and production company based in Thailand) applies AI on the frontline, transforming emergency response from reactive troubleshooting to proactive, guided intervention. Chang-Prompt, an AI-powered platform that transforms crisis management by acting like an ICU triage—enabling immediate action and diagnosis. You’ll hear real business use cases, learn how AI enables faster decisions, and understand the challenges.

With the EU AI Act in place, achieving presumption of conformity for high-risk AI systems requires a clear strategy. This session provides a concrete overview of the current state of standardization and key harmonization efforts. Frank will examine how to apply essential standards for AI management systems, risk assessment frameworks, and conformity assessment processes. Participants will learn how to align their AI system with emerging regulatory expectations and efficiently navigate compliance.

AI has been hailed a breakthrough just as often as a passing fad, but with ChatGPT it looks like it is here to stay. But does this also hold true for AI in forecasting and Predictive Anlytics? With AI the go to forecasting algorithm in Electricity Demand, Computer Server Load and Ride Sharing, it is slow to catch on in other areas. Sven will share personal case studies from Beiersdorf, Hapag-LLoyd, ABInBev etc, travelling the hype cycle twice from 2004 towards the next disillusionment post 2025

Shaping Europe’s digital future, ethics play a decisive role in implementing the AI Act. Good Data Governance aims to achieve both, efficient data management and Responsible AI. Frank and Susanna present their federated model, where data stewards collaborate with so-called data domain owners and data scientists. Using their machine-learning-based salary-prediction tool as a case study, Frank and Susanna demonstrate how fairness can be embedded into applications, highlighting how they address challenges such as gender disparities.

Zeiss developed GraphChat, a package that is being used by an increasing number of projects within Zeiss. GraphChat takes a natural language question, leverages a knowledge graph and LLM and comes back with an answer. This session addresses the limitations of traditional chatbots by enhancing user engagement and information retrieval accuracy through proactive dialogue and advanced architecture. Learn about the Planner and Executor Multi-Agent architecture, the role of Knowledge Graphs in improving context-awareness and the capabilities of GraphChat in executing precise queries.

This tech deep dive will take you from zero to a level of understanding of Agentic AI, multi-agent collaboration and Model Context Protocol (MCP) that enables you to start building own use-cases right away. Tomas will start with the simplest examples and ends up with advanced multi-agent collaboration patterns such as SWARM (compete), and GRAPH (collaborate) utilizing open source standards and libraries such as MCP, and Strands Agents SDK.

The assumption, that modern, complex models automatically outperform older, simpler ones, still permeates many organisations and decision-makers. Niklas will show analytically (using Bias-Variance Trade-Off and the interpolation threshold) when this holds true and when simple, transparent models perform as well or better than the most sophisticated modern AI tools. He will showcase how he used this approach to design a new, fully transparent, high-performance credit score with the Schufa Holding AG.

Introduction of an AI assistant for 3,000 employees at Hermes (Otto Group). 75% of users save over 30 minutes per week through this assistant. Questions addressed during the lecture: how do I successfully implement an AI assistant? How do I communicate and train? What application cases are available? What should be considered in terms of data protection and compliance? AI assistant similar to M365 Copilot and ChatGPT. Experienced speaker, e.g., Swiss Online Marketing & IFS Ultimo Customer Day. 

Querying specialized data like parliamentary records is crucial. This session introduces FELD M’s RAG Blueprint, an open-source framework for building production-ready RAG systems. João will showcase its application at the example of German Bundestag speeches. You’ll learn how to develop scalable and observable RAG solutions. Key takeaways include a practical path to deploying robust RAG for complex Q&A and the value of a comprehensive, production-ready framework.

Graph structures effectively model complex relational data across many domains. Tudor and Jia demonstrate the application of Graph Neural Networks (GNNs) on fraud detection in Trustpilot by modeling the intricate relationships between businesses and consumers. Their work showcases how GNNs can capture the complex patterns of fraudulent activities that traditional methods often miss. Additionally, they discuss model explainability, ensuring that the learnings of our GNN align with domain expertise.

Machine Learning adds complexity to product design—not just in implementation, but even more during product discovery, where user value is explored. At edding, the easycheck team transformed to an empowered product team, collaborating across disciplines by rethinking roles, rituals, and structures. This session shares insights on making ML research and implementation a first-class citizen and enabling product managers and designers to work with uncertain model outputs.

Agentic approach represents the future of AI application development. Learn effective methods for creating AI agents that utilize both proprietary and open source LLMs, accompanied by practical demonstrations showing these agents collaborating to solve real industry challenges. The tech showcase will feature n8n as a low-code AI agents platform alongside Agno and CrewAI frameworks for Python-based AI agents. Also learn about vendor and custom built MCP servers & integrate it with AI Agents.

Accurately forecasting expenditures is crucial for budgeting and policy planning. This session presents a multi-task LSTM model designed to predict annual federal student aid (BAföG) expenditures for the current and next two years. Using 16 state-level time series, it addresses extreme intra-year volatility, strong seasonality, and the risk of overfitting with small datasets. Attendees will gain insights into overcoming these challenges and improving real-world time series forecasting models.

This talk will highlight how the electric car manufacturer NIO leverages machine learning to develop NOMI, the world’s first in-vehicle artificial intelligence. The session will share lessons learned from integrating LLMs into the existing voice assistant architecture, including the challenges and opportunities this integration has presented for both the product and the organization. 

Post-training model compression is a powerful way to make Large Language Models accessible in resource-constraint environments. Oftentimes, however, practitioners find themselves adjusting their needs to a set of predefined compression parameters, rather than the other way round. In this talk, I will advocate for model compression that aims to be as simple as image compression: enable users to choose a model size that fits their own needs.

The Table Captains present the three key take-aways from their discussion round.