Agenda Machine Learning Week Europe 2026

On paper, we have never been more data-driven. But the day-to-day culture of many organizations—the habits, the gut instincts, the questions leaders think to ask—often struggles to keep pace. The key to true business transformation is ensuring that decision-making actually matches the power of the technology behind it. In this keynote, based on his book Data Inspired, Sebastian Wernicke reveals how to bridge the gap between brilliant technology and daily operations.

Building a user-facing GenAI product within the fast-paced news environment of a public broadcaster is hard. This talk shares a collection of hard-earned architectural lessons from a real use case: from Denial-of-Wallet prevention, to why caching is key for controlling costs. We discuss concrete engineering solutions for preventing hallucinations or dealing with interconnected documents. Join this session to bypass the trial-and-error phase and discover what it really takes to run GenAI.

Everyone talks about jailbreaks and content filters but what stops LLM abuse when the filter fails? An attacker can make every single request look harmless while the overall pattern is clearly hostile. Drawing on a decade of hunting threat actors in cybersecurity, this session moves beyond prompt-level safety to behavioural, actor-level detection: spotting who is misusing a system, not just which prompt is bad. You will leave with a practical, layered way to think about AI safety.

Remember when AI-generated content sounded like a robot reading a script? Those days are over. Automating editorial content usually means choosing between scale and trust, at wetter.com we needed both. I will showcase the pipeline behind our multilingual forecast videos and weather articles, from data to narrated video, where LLMs, TTS and meteorologists each fit. Learn how far AI quality has come, and take home a blueprint for scaling content without losing accuracy or trust.

Most companies have data spread across databases, APIs, and reports. Answering a real business question means pulling from all of them at once. This session shows how AI agent patterns – Planner, Aggregator, and Verifier and turn that chaos into reliable answers. I will build a working demo live. Participants learn how to design agent systems, not just prompts. Key questions: When should humans review agent output? How do you handle conflicting sources? What makes an agent production-ready?

PAYBACK’s integrated search system combines hybrid search with ‘Pointy,’ our AI chatbot, enabling users to discover partners, coupons, and rewards, or receive instant answers to their questions across the entire PAYBACK ecosystem. This deep integration within our platform showcases an innovative application of AI in e-commerce. Our presentation will cover both the development of a highly scalable AI agent and the engineering loop that ensures its reliability and improvement in production.

Forecast the Bundesliga ranking this season, and the outcome of every match. But no black boxes allowed: here’s how to model under football-level uncertainty, honestly and traceably. You’ll learn to predict matchdays with a model whose every prediction traces back to its drivers, calibrate its confidence, and watch it drift as form, injuries and transfers play out. (This same discipline that keeps a model honest in football keeps it trustworthy once it leaves demo stage at your company.)

When REWE launched its new loyalty programme, one long-standing data product required a significant update: lifestage customer segmentation. We’ll show how we replaced a legacy linear-regression system prone to drift with a robust model. It scales survey ground truth from fewer than 20,000 respondents to millions of weekly transactions, mitigating data bias while maintaining prediction stability without frequent retraining.

Drug discovery, new materials, and manufacturing planning look like different worlds — but from a data perspective they pose the same problem: find good candidates in a design space far too large to search. And what counts as “good” is yours to define — a molecule that’s easier to make, a material that’s cheaper, a schedule with less waste. We’ll walk through the idea of GFlowNets, demo it live in code, and show how you can adapt to your own work.

Why it matters: In banking, churn must be identified at scale across multiple products, under strict regulatory constraints, using interpretable models despite severe class imbalance. What we cover: how an ML factory standardizes and accelerates churn deployment while ensuring explainability and governance. Takeaways: how banks can reduce time‑to‑production and deliver controlled, measurable business impact with ML.

From the start, Machine Learning Week has been the place to discuss and share our common problems. These are your people – they understand your situation. Often rated the best part of the event, sharing your problems with like-minded professionals is your path to answers, a little empathy, and a stronger professional network. Table Discussion Topics will be announced soon. Come prepared with your problem you are solving, or a question that needs answering. And then, be ready and willing to help others in this, their hour of need.

1. How to get AI into production – Ari Joury

95% of AI pilots delivered little or no business value in 2025 (MIT NANDA). Technical issues are consistently reported as the easiest part

1. Why is AI so hard to prove and scale right now?
2. What strategies have you tried to make your AI projects successful?
3. How will AI’s business value evolve in the next 3 years, realistically?

2. Machine Learning from Data You Can’t Move – Anthonette Ochieze

1. Where has your team been blocked from using data because of privacy rules, contracts or regulation, and what did you do instead?
2. Federated learning, synthetic data, differential privacy and anonymisation all promise a way through. Which have you tried in production, and where did they fall short?
3. How do you get legal, compliance or data owners comfortable enough to sign off on a privacy-preserving approach?

From a law firm’s perspective, preparing a company for an IPO means bringing together client documents, web research, communications, regulations and compliance requirements. This session focuses on one of the key sections of an IPO filing: Risk Factors. We will show how premium financial market data, bespoke data curation, AI agents, classical data science, NLP and statistics are leveraged to draft and evaluate this section, and what it takes to use AI as part of a complete, reliable solution.

From the start, Machine Learning Week has been the place to discuss and share our common problems. These are your people – they understand your situation. Often rated the best part of the event, sharing your problems with like-minded professionals is your path to answers, a little empathy, and a stronger professional network. Table Discussion Topics will be announced soon. Come prepared with your problem you are solving, or a question that needs answering. And then, be ready and willing to help others in this, their hour of need.

AI projects rarely fail on technology or data, but skepticism. Matthias Göbel shows how to build trust where mistrust reigns and facts alone fall short. He shows real word examples concrete principles for turning skepticism into acceptance. Your last AI project that stalled: who didn’t come along, and why? Which skepticism toward your AI work is actually justified — and what does it teach you? What would make you not just sound more convincing, but act more trustworthy?

Companies consider AI-systems to replace forecasting in legacy systems, hoping for improved accuracy and automation, even abandoning human demand planning teams.These decisions are often based purely on technological hype. We present a case study of the cost savings achievable by 1% Forecast Accuracy from AI forecasting (depending on data) and automation, against the cons of lost accuracy from judgmental adjustments showing that AI can improve Analytics but not repace human experts.

Generative AI is changing how customers discover, compare, and choose car dealers. In this practical workshop, participants will learn what makes a dealership visible in AI-driven search, from inventory data and local relevance to reviews, service information, and clear website content. They will leave with a simple framework to assess their current visibility, identify gaps, and prioritize actions that can improve discovery, trust, and lead generation.

Reliable waiting time predictions are essential for customer satisfaction during roadside assistance, yet difficult to achieve in dynamic operations. In this session, we show how we partnered with ÖAMTC to develop machine learning models that significantly improved prediction accuracy over the existing approach. We share practical lessons on choosing the right models, addressing data limitations, and continuously improving predictions through operational feedback and new data.

Anonymisation does not work. A few quasi-identifiers can unmask people in an anonymous dataset, and pooling sensitive data centrally builds a target. So why move the data at all? This session works through federated learning, differential privacy and synthetic data, with live code. For each, what it protects, what it costs, where it breaks. It draws on my published architecture, which cut data exposure by 99.2 per cent. You will leave able to pick the right method for your threat model.

How does Miele create customer value through AI? This talk shares key lessons learned on the journey from initial AI applications to scaling AI across a portfolio of connected appliances. It sheds light on strategic as well as technical aspects behind smart sensing, machine vision and LLM-based applications at Miele. The presentation highlights critical capabilities to deliver AI products for IoT appliances and sketches how large foundation models impact the innovation process itself.

In 20 minutes, two participants will each present their current project and their biggest challenge. Together, they will look for new solutions

How do you use ML to really understand your manufacturing process? Complex processes are hard to model – everything affects everything, and standard models struggle to capture that full chain of effects. In this talk, we show how CeramTec and Merantix Momentum tackled this at the ML level: After transitioning the shop floor to be data-first, we used Explainable Graphs to predict outcomes mid-process and give decision-makers a granular view of what drives what across the line.

Old-school machine learning, transformers or LLM agents: which one detects fake news best? In this 60-minute hands-on competition you build a fake news classifier with the tools you trust most, with or without an AI coding assistant. Two leaderboards, one for accuracy and one for quality, show which approach wins. Bring a laptop and pick your side.

One dataset. One hour. Three ways to win.

Can a carefully tuned TF-IDF model still beat a transformer? Can an LLM agent beat them both? In the Tool Battle you find out live. You get a set of real and fake news articles, 60 minutes and a free choice of weapons. Pick your side:

• The classics: traditional machine learning. TF-IDF, SVD, logistic regression, Light-GBM. Fast, lean and proven. Show that solid feature engineering is far from retired.
• The deep learners: transformers. DistilBERT, RoBERTa and friends, used as embeddings or fine-tuned on a free Colab GPU.
• The new kids: agents and LLM APIs. Zero-shot and few-shot prompts, agentic pipelines, GPT, Claude or Gemini as the classifier. No training loop required.

Or mix them: hybrids of all three are welcome. Write every line yourself or let ChatGPT, Claude Code, Cursor or Codex write it for you. Vibe coding is explicitly allowed, and so is twenty years of ML experience. At the end, two live leaderboards settle the question. One ranks the solutions by ROC-AUC on hidden test articles. The other ranks them by quality: an LLM agent, supervised by experts, reviews every notebook for correctness, data leakage and soundness of the approach. So the highest score is not the only way to the top, and a clean, simple model can beat a flashy one. Whether you belong to the “old guard” of machine learning or to the “new kids” of generative AI: bring a laptop, pick your tools and prove that your way works best. All you need is basic Python and a Google account for Colab.

Balancing data sovereignty with AI access is a critical challenge. Join Stadtwerke Norderstedt and synvert to explore our 4-year journey scaling from a handful of GPUs to a multi-cluster Kubernetes architecture. We will share how moving workloads off the cloud enabled a secure AI ecosystem with on-prem Chat, RAG, and agentic workflows. Attendees will learn to build sovereign AI that users love, gaining insights into adoption metrics, performance benchmarks, and hard-learned pitfalls.

Most Agentic AI initiatives never make it beyond impressive demos. This interactive clinic uncovers the technical, architectural, and organisational roadblocks that prevent production success. Through real-world examples and practical diagnostics, participants will learn how to identify failure patterns, design production-ready AI systems, and apply proven principles to build reliable, scalable, and trustworthy Agentic AI.

Machine learning evolved as an attempt to make computers – imagined as perfect machines – understand the imperfect and noisy data produced in the real world. But large computer systems are far from perfect, and modern software systems produce data as imperfect as any real-world system. We discuss how we build tools that help understand, and prevent, failures in large software systems, using methodologies ranging from Bayesian modeling over Causal Analysis to Transformer-based neural networks.

The model with the best statistical score may not deliver the best business result. Bring a machine learning project—and the metric you currently use—to this interactive clinic. Using a live Python simulation dashboard, Dean Abbott will translate selected audience scenarios into simulated populations and model predictions. Together, we’ll explore how outcome distributions, uneven prediction errors, and business costs change model rankings—and quantify the opportunity cost of choosing models by the wrong metric under those assumptions. A real-world fundraising example will ground the discussion in actual data. Participants will leave with practical evaluation ideas and Python code they can adapt to their own projects. No coding, software installation, or data sharing is required during the session.