2.0 -//Pentabarf//Schedule//EN PUBLISH LZRVAL@@cfp.pydata.org -LZRVAL Opening en en 20251209T090000 20251209T091000 001000

Opening

Welcome to PyData Eindhoven 2025 PUBLIC CONFIRMED Opening https://cfp.pydata.org/pydata-eindhoven-2025/talk/LZRVAL/ Auditorium PUBLISH R7NMXB@@cfp.pydata.org -R7NMXB Opening Keynote by Maurits Hendriks: “The Science of Showing Up”: Why Confrontation, Selection, and Data Matter in the Journey to Elite Performance en en 20251209T091000 20251209T095500 004500

Opening Keynote by Maurits Hendriks: “The Science of Showing Up”: Why Confrontation, Selection, and Data Matter in the Journey to Elite Performance

Drawing on decades in Olympic and professional high-performance environments, this keynote reveals how the best athletes and teams design structures to peak at exactly the right moment. Ultimately, the science of showing up is the art of combining human resilience with intelligent systems—being ready, physically, mentally, and organizationally, when the world is watching. PUBLIC CONFIRMED Keynote https://cfp.pydata.org/pydata-eindhoven-2025/talk/R7NMXB/ Auditorium PUBLISH XMJ9LX@@cfp.pydata.org -XMJ9LX Scaling Retail Planning at IKEA: Orchestrating Sales, Fulfillment and Capacity Assessment with Metaflow en en 20251209T100000 20251209T103000 003000

Scaling Retail Planning at IKEA: Orchestrating Sales, Fulfillment and Capacity Assessment with Metaflow

Retail planning at IKEA is more than just predicting sales, it is about connecting various forecasts that inform and depend on one another: Sales forecasting: How much will customers buy? Fulfillment planning: Can we ensure availability? Capacity assessment: Do our stores and distribution centers have enough capacity to handle the volume? Each of these domains has its own models and ways of working. Previously, these teams built and deployed everything independently. Making it difficult to align predictions and maintain visibility across the retail chain. To improve this process, IKEA adopted Metaflow as a data science platform. Metaflow provides a framework for developing, running, and monitoring pipelines. The patterns and solutions shared during the presentation can apply to any organization dealing with (complex) workflows, interconnected predictions, or resource/scaling challenges. Target audience: Data scientists, data engineers, ML practitioners, and technical leads interested in: Workflow orchestration and reproducibility. Cross-team collaboration in data science. Scalable forecasting systems in production environments. Background knowledge: Basic understanding of data pipelines and workflow orchestration. No Metaflow experience required. Talk outline: 0-10 minutes: Introduction to IKEA’s Retail Planning Problems that you encounter coordinating these processes in 30+ countries Why traditional pipelines did not succeed 10-15 minutes: What we needed: A system that could handle complex dependencies, scale and let non-engineers contribute safe and easy Introduce Metaflow as an OS framework solution 15-25 minutes: Quick introduction how Metaflow works in practice Show how Metaflow handles the Retail Planning chain Show how Metaflow works together with Argo Workflows Show how Metaflow helps us scale and manage our resources at the same time 25-30 minutes: recap and questions PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/XMJ9LX/ Auditorium Yannick Mariman Harlley de Lima PUBLISH NHFJJW@@cfp.pydata.org -NHFJJW AI-Powered Web Scraping: From Data Collection to Strategic Insights en en 20251209T104500 20251209T111500 003000

AI-Powered Web Scraping: From Data Collection to Strategic Insights

Collecting web data is getting harder—between messy datasets and stronger bot defenses, traditional scraping often falls short. This talk shows how combining Python tools (Scrapy, Playwright, Selenium) with AI/ML can turn raw, unstructured data into clear insights. We’ll explore practical best practices, real-world case studies, and how SMEs use sentiment analysis pipelines to make smarter marketing and product decisions. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/NHFJJW/ Auditorium Yevhenii PUBLISH LURJEK@@cfp.pydata.org -LURJEK Efficient Time-Series Forecasting with Thousands of Local Models on Databricks en en 20251209T112000 20251209T115000 003000

Efficient Time-Series Forecasting with Thousands of Local Models on Databricks

Industries like energy, retail, and logistics often face a critical problem: forecasting demand or consumption for thousands of entities, each with unique patterns. Global models may miss local nuances, while training individual models can overwhelm traditional pipelines. In this talk, we’ll explore a practical and scalable solution built on Databricks—using pandas API on spark to train and predict thousands of local models in parallel. You’ll discover how to: - Train and serialize ML models per group efficiently. - Use binary model storage to overcome MLflow’s RPS limits. - Generate and forecast future data asynchronously at scale. We’ll discuss the trade-offs between MLflow Registry, Unity Catalog, and inline model execution, and how this approach powers forecasting for thousands of groups in the energy sector. Whether you’re a data scientist or ML engineer, you’ll leave with actionable ideas for scaling your own forecasting workflows on Databricks PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/LURJEK/ Auditorium Daria Mustafina PUBLISH 9FRSYC@@cfp.pydata.org -9FRSYC Finding trash in waste en en 20251209T125000 20251209T132000 003000

Finding trash in waste

After waste collection of lightweight packaging waste, in the Netherlands known as PMD (plastic-, metaal- en drankkartons), the incoming material is visually inspected by an operator to extract contaminants that disturb the sorting and recycling. Typical contaminants are non-packaging material and large rigids and foils. At the applied research institute NTCP, we investigated how to automate the detection process in order to have an objective method to inspect a PMD waste stream. This improves consistency and thereby quality and reduces required time needed for a thorough assessment. We decided to use cameras in a top view of the pile of items, as the inspection is performed. For every item, we first need to locate it in the image and segment the pixels of it. We built a data pipeline using vision foundation models. To determine whether an item is considered a contaminant or PMD packaging, we found that the class of contaminants had a too broad variety to build a classification network. Therefore, we chose to explore the feasibility of anomaly detection models. Several dedicated anomaly detection models were trained. Finally, we learned that using a combination of foundation models for both visual as well textual features, resulted in a suitable distribution to distinguish contaminants from PMD packaging. With a prototype set-up, we have showcased a proof-of-principle of a well working and promising automatic detection system on actual waste, given the variety in both PMD and contaminants. images, given the variety in both valid waste and unwanted items. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/9FRSYC/ Auditorium Tom Koopen PUBLISH RUTXQ3@@cfp.pydata.org -RUTXQ3 From €1M License to In-House Success: How We Built a Real-Time Recommendation System and Saved Millions Doing It en en 20251209T132500 20251209T135500 003000

From €1M License to In-House Success: How We Built a Real-Time Recommendation System and Saved Millions Doing It

We started our journey of developing an in-house personalized banner recommendation system at Bol, by replacing an expensive third-party tool. We are going to talk about our decisions from ML to explore-exploit balance of our recommendations, sharing our story of building the system from scratch. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/RUTXQ3/ Auditorium ALI KOHAN PUBLISH JBPQS9@@cfp.pydata.org -JBPQS9 Federated Data, Centralized Action: A Governance Model Powered by Lakewatch and Databricks System Tables en en 20251209T141000 20251209T144000 003000

Federated Data, Centralized Action: A Governance Model Powered by Lakewatch and Databricks System Tables

Learn how this approach drives centralized action, empowers domain teams, and eliminates governance bottlenecks while optimizing Databricks-related costs. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/JBPQS9/ Auditorium Cristiano Cortez Da Rocha Frank Chidi Mbonu PUBLISH TN83BD@@cfp.pydata.org -TN83BD Building, Deploying and Managing AI Agents at Scale en en 20251209T144500 20251209T151500 003000

Building, Deploying and Managing AI Agents at Scale

The hype around Agentic AI is real, but so are the challenges. Creating a single agent is straightforward.. But how do you scale effectively? If you don't do it properly, critical issues will arise: spiraling costs, fragmented oversight, and major governance risks. Your agents need to be governed by a set of fundamental design principles. In other words, you need a platform. But what does a “platform” truly mean in practice? Sander will provide a pragmatic deep dive—no hype, no superficial demos—into the key pillars of this foundational platform: - Infrastructure & Serving: reliable serving patterns with Vertex AI Agent Engine. - State & Memory: Implementing session management and long-term memory for personalized interactions using Google Cloud's MemoryBank. - Orchestration: Using frameworks like Google's Agent-Development-Kit (ADK) - Integration: We will take a look into the different options to connect your agents with external services and data. - Security: Highlighting current authentication and authorization challenges for autonomous agents. - Governance & Cost Control: Establishing robust guardrails for cost, security, and compliance from day one. - Observability & Evaluation: Implementing rigorous monitoring and evaluation to prove business value and ensure reliability. This talk offers a pragmatic, engineering-focused approach to building a resilient, manageable, and impactful AI agent ecosystem on Google Cloud. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/TN83BD/ Auditorium Sander van Donkelaar PUBLISH 8YNNH9@@cfp.pydata.org -8YNNH9 Responsible Human-Agent Ecosystem: Making Mission Critical Systems Situationally Aware en en 20251209T152000 20251209T155000 003000

Responsible Human-Agent Ecosystem: Making Mission Critical Systems Situationally Aware

Using technologies such as 3D digital twins, computer vision, and predictive ML models, our systems create a live, interactive view of the internal working of a large, multi-system operational environment. AI agents can detect anomalies, track threats, reason about context, and support operators with transparent, trustworthy insights. For developers and technologists, this work sits at the intersection of AI, geospatial engineering, high-performance computing, and human-centered design offering a chance to build systems that truly matter. Join us to see how cutting-edge tech is shaping the next generation of situationally aware capabilities. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/8YNNH9/ Auditorium Ipsit Dash Bart-Peter Smit PUBLISH XTSZYZ@@cfp.pydata.org -XTSZYZ Hier is Vincent. Eén en al oor. en en 20251209T155500 20251209T162500 003000

Hier is Vincent. Eén en al oor.

Can the public engage directly with Vincent van Gogh? Based on more than 800 letters in which he describes his circumstances, work and emotions, a so-called Digital Twin has been developed. Enriched with additional sources and skills, and created with respect for copyrights, ethics and privacy. This gives the public the opportunity to experience his thoughts, struggles, vulnerability, and passion in a vivid and meaningful encounter. PUBLIC CONFIRMED Keynote https://cfp.pydata.org/pydata-eindhoven-2025/talk/XTSZYZ/ Auditorium Rob Mulder PUBLISH PHUKC8@@cfp.pydata.org -PHUKC8 Closing en en 20251209T162500 20251209T164000 001500

Closing

Closing PUBLIC CONFIRMED Closing https://cfp.pydata.org/pydata-eindhoven-2025/talk/PHUKC8/ Auditorium PUBLISH 7GEDFP@@cfp.pydata.org -7GEDFP Developing a Nation-Wide Padel Rating System: A Data-Driven Approach en en 20251209T100000 20251209T103000 003000

Developing a Nation-Wide Padel Rating System: A Data-Driven Approach

Elo systems are widely used to estimate player levels in situations where everyone plays against different opponents. The principle is simple: after each match, your rating shifts based on the expected win probability. Beating a stronger team increases your rating more than beating a weaker one. In the Netherlands, padel initially used the existing tennis rating system. But tennis and padel have different characteristics that make this challenging. In tennis, a rating gap often translates quite predictably into a win probability. In padel, more factors come into play: mixed-gender teams, asymmetric pairings (e.g., two men versus a man and a woman), and the sport’s specific match dynamics. Developing an effective padel rating system is a real-world problem. We found that data analysis is crucial, but without sport-specific knowledge and context, conclusions can be misleading. We analyzed data from more than 300,000 padel matches and tested a range of models—from standard Elo variants to fully custom-built algorithms. Each system was evaluated on, amongst other things, two core metrics: speed (how quickly a system converges to a realistic player level) and efficiency (how stable ratings remain without unnecessary fluctuations). This presentation will walk through the technical process, the model comparisons, and the padel-specific decisions that led to a rating system designed to work in real competition. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/7GEDFP/ Ernst-Curie Max Brouwer PUBLISH 38RZBP@@cfp.pydata.org -38RZBP Identifying playstyles in football through spatial networks en en 20251209T112000 20251209T115000 003000

Identifying playstyles in football through spatial networks

Football analysts and coaches rely on manual video analysis to identify playstyles and tactical behaviors, a process that is time-consuming and often subjective. This talk introduces an innovative approach that leverages tracking and event data to automatically identify playstyles in the critical moments leading up to a shot on goal. By employing advanced network science and football analytics, this method offers a comprehensive and objective analysis of both attacking and defensive strategies. In this presentation, we will start with the introduction of spatial networks, where we approach football players as an undirected graph. Then, we will look into the developed spatial network metrics that are used in the identification of playstyles. These are summarized into three groups, measuring space control, pressure, and connectivity. With these metrics, we can use clustering techniques to obtain different attacking and defending playstyles. We conclude with the results of the clustering techniques and the possible interpretation and applications of the results. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/38RZBP/ Ernst-Curie Annemarijn Blom PUBLISH 9KFFYT@@cfp.pydata.org -9KFFYT FootballBERT: Encoding player identity in vectors with Transformers. en en 20251209T125000 20251209T132000 003000

FootballBERT: Encoding player identity in vectors with Transformers.

In football analytics, player identity is still often encoded using one-hot vectors or individual statistics — highly volatile over short time periods. Such approaches ignore the relational context between teammates and opponents, and coaches tactical demands. FootballBERT changes that paradigm. Inspired by NLP breakthroughs, it applies the Transformer architecture to lineup data, learning contextual player embeddings through a Masked Player Prediction (MPP) objective. Each embedding encodes a player’s identity through patterns of who they play with, against, and under which tactical setups. In this talk, I’ll walk through: - Why embedding player identity into dense vectors matters for football analytics. - How FootballBERT is trained end-to-end from raw lineup data and positional features across 170K+ matches. - Empirical insights on what these embeddings capture — especially their ability to generalize across leagues. - How such representations enable downstream applications like BALLER-Transfer Portal, an AI model that predicts how any player would perform in any tactical context with unprecedented granularity. This session bridges deep learning, transformers, and sports data science — showing how domain-specific foundation models like FootballBERT can power the next generation of context-aware analytics. Whether you’re an ML researcher, AI engineer, or football data scientist, you’ll leave with a concrete understanding of how to build, train, and apply transformer models beyond text — starting from real-world, messy data. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/9KFFYT/ Ernst-Curie Achraff ADJILEYE PUBLISH AHXR33@@cfp.pydata.org -AHXR33 Football is complex, but your code doesn’t have to be — meet DataBallPy and a practical deep dive into pressing en en 20251209T132500 20251209T135500 003000

Football is complex, but your code doesn’t have to be — meet DataBallPy and a practical deep dive into pressing

Answering complex data-driven questions about football tactics is a lot of fun. Trying to parse football data, synchronize tracking and event data, and determine which individual player actually has ball possession is just tedious work, which limits the time to actually answer the questions you wanted to answer in the first place. DataBallPy is an open-source Python package designed to abstract away these repetitive tasks, allowing analysts and developers to focus on building models and answering their questions. In the first half of this talk, we introduce DataBallPy’s architecture and core functionality (and how it differs from existing packages like Kloppy). Using code examples, we will show how you can perform multiple preprocessing steps and visualise the data, and we will showcase some of the implemented features. Equally important, we value transparency and education. Therefore, we have some elaborate documentation on how all features are implemented. In short, DataBallPy will quickstart your process in answering your questions. The second half of the talk presents a practical case study: how the Royal Belgian Football Federation (RBFA) uses open-source packages, among which DataBallPy, to analyse the efficiency and effectiveness of pressuring the opponent. The RBFA was inspired by the Common Data Format (CDF) (Anzer et al., 2025) to document and store data. Similarly, DataBallPy functionalities were made to work with the CDF as intended. The research builds on Bekkers’ pressing model, which uses the Time To Intercept (TTI) concept to estimate how quickly a defender can reach a specific target location on the pitch, allowing it to represent pressure. During the talk, we will showcase how the RBFA got to answer the following two questions: (1) Does external load differ significantly between successful and unsuccessful pressing actions? and (2) How does external load vary across pressing actions initiated from high-, mid-, and low-block defensive positions? PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/AHXR33/ Ernst-Curie Alexander Oonk Tygo Nikamp PUBLISH GGZ9TA@@cfp.pydata.org -GGZ9TA Planning Hockey Careers With Python en en 20251209T141000 20251209T144000 003000

Planning Hockey Careers With Python

## Objective To share the process of building a data-driven career path planner for young hockey players, highlighting both the technical challenges and the broader impact of applying data science in real-world decision-making. ## Outline - Why hockey (and why career planning matters in sports) - Data collection & wrangling: turning messy, incomplete performance records into usable data - Modeling progression: techniques to identify likely career trajectories and milestones - Communicating results: designing tools that resonate with non-technical stakeholders (players, coaches, parents) - Lessons learned: challenges, surprises, and what translates to other domains beyond sports ## Central Thesis Data science can provide valuable guidance in high-stakes, personal decision-making — but building tools for non-technical users requires more than just models. It demands thoughtful data cleaning, careful feature selection, and communication strategies that make insights accessible and actionable. ## Audience This talk is for data scientists, analysts, and practitioners who are curious about applying machine learning to sports. It will appeal both to those interested in the technical process (data wrangling, predictive modeling) and those interested in building data products that impact real people. ## Tone / Type of Talk Informative and light-hearted — grounded in practical data challenges, with a touch of storytelling from the hockey world. This is not a heavily mathematical talk at all, suitable for broader audience as well. ## Key Takeaways - How to wrangle and model messy sports data for career planning use cases. - How to design predictive models that communicate uncertainty and possible paths. - Lessons on building data tools for non-technical stakeholders. - Inspiration for applying data science to creative and impactful real-world problems. ## Background Knowledge Expected No prior knowledge of hockey is required — we’ll keep the sports side light and fun. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/GGZ9TA/ Ernst-Curie Jaroslav Bezdek PUBLISH EGJV7E@@cfp.pydata.org -EGJV7E Optimizing fantasy basketball decisions with Python: linear & integer programming for roster management en en 20251209T144500 20251209T151500 003000

Optimizing fantasy basketball decisions with Python: linear & integer programming for roster management

Core Content: This talk teaches Linear and Integer Programming fundamentals through fantasy basketball examples, demonstrating when to use each approach for optimization problems. Session Structure: -Linear programming fundamentals - Integer programming fundamentals - Demo: start/sit optimization - Demo: free agency optimization - Real life applications I will also explain rules of different fantasy basketball formats. Prior Knowledge Expected: Python (no basketball knowledge needed but not hating on sports will help) PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/EGJV7E/ Ernst-Curie Pawel Kapuscinski PUBLISH K9LDVT@@cfp.pydata.org -K9LDVT xReceiver: a GNN approach to the evaluation of the decision-making process of passing options in football en en 20251209T152000 20251209T155000 003000

xReceiver: a GNN approach to the evaluation of the decision-making process of passing options in football

During my internship at the Royal Belgian FA, my supervisor received a request from the performance analyst: how can we analyse the quality of passes using AI? We generally consider a player to be a good passer if most of their passes are successful, but sometimes passing the ball to that player is not the right option if there is someone else who is less marked and in a better position! Can players really identify the right teammate in the right situation? Outline of what will be discussed during the talk: 1. Introduction (2–3 minutes). 2. From data to graph through synchronization (5 minutes). 3. Graph Neural Networks (5 minutes) 4. The xReceiver model (7-8 minutes) 5. Applications in Opponent Analysis and Scouting (5 minutes) Questions and feedback will be welcomed at the end, since these can really help with the improvements of this project. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/K9LDVT/ Ernst-Curie Gabriel Masella PUBLISH 9KLLZY@@cfp.pydata.org -9KLLZY Beyond One Model: Scaling, Orchestrating & Monitoring en en 20251209T100000 20251209T103000 003000

Beyond One Model: Scaling, Orchestrating & Monitoring

# Description Building a single machine learning model is one challenge. Running thousands of them in production — reliably, efficiently, and transparently — is another. As Vanderlande adopted its use of anomaly detection, we faced the reality that success wasn’t about the accuracy of one model, but about the MLOps infrastructure to scale, orchestrate, and monitor many. We are two data scientists who gradually found ourselves tackling problems that look a lot more like ML engineering. In this session, we’ll share our journey in a practical and informal way, showing the real steps we took, the hurdles we hit, and the solutions we built. We’ll walk through the orchestration layer that coordinates training and deployment for thousands of models, highlight practical solutions for scaling pipelines, and discuss how we keep the system resilient when failures occur. A short code snippet will illustrate how lightweight the deployment process can be, and we’ll close with a look at the monitoring framework that provides visibility and reliability across the entire fleet. Whether you are just starting to scale your ML initiatives or already struggling with operational complexity, this talk will offer lessons learned and concrete practices for moving beyond one model. # Outline **1. Vanderlande’s use case (5 min)** - Vanderlande specific use case - Anomaly detection at scale - Why *beyond one model* became a necessity **2. Machine learning orchestration & pipeline (13 min)** - Intro: One model ML Architecture - Architecture & orchestration design for multiple models - Isolated pipelines to avoid cascading failures **3. Code snippet & peek into monitoring (7 min)** - Inference orchestrator code example - CI/CD deployment pipeline - Monitoring framework: jobs, models, anomalies at scale - Lessons learned & what we’d do differently **4. Wrap-up & Q&A (5 min)** - Key takeaways - Quick audience questions PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/9KLLZY/ Planck-Bohr Azucena Morales Vi Chu PUBLISH EMUNDT@@cfp.pydata.org -EMUNDT From Data Lake Entanglement to Data Mesh Decoupling: Scaling a Self-Service Data Platform en en 20251209T112000 20251209T115000 003000

From Data Lake Entanglement to Data Mesh Decoupling: Scaling a Self-Service Data Platform

1. What exactly is architectural decoupling 2. Identify how hidden coupling creeps into data lakes and data meshes. 3. Learn technical design patterns for decoupling ingestion, transformation, and ownership using open tools. 4. Understand how to scale self-service data mesh principles without breaking governance or lineage integrity. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/EMUNDT/ Planck-Bohr Geert Jongen PUBLISH J3ECG7@@cfp.pydata.org -J3ECG7 Scaling Python to thousands of nodes with Ray en en 20251209T125000 20251209T132000 003000

Scaling Python to thousands of nodes with Ray

DESCRIPTION I don’t read as much as I used to. As a result, ploughing through the Wheel of Time series by Robert Jordan has become a multi-year project for me. With 2787 named characters across 14 books, I often have to search on wikis for what happened six books ago. And who was that one minor character again…? With AI we can do better. What if I fine-tuned an LLM with RAG to help me keep track of the entire saga? I could ask questions and it would give me spoiler-free answers… ✨
 Whipping up a few lines of Python to train and tune models is easier than ever. But while executing scripts on a laptop or VM is trivial, it becomes much more difficult when you want to parallellize such a workload across an entire cluster. That’s where Ray comes in. With a few lines of code, we can modify our code to distribute our model training to tens or hundreds of machines. If you ever wondered how you companies like OpenAI train foundational models: this is how.
 In this talk, I’ll cover Ray’s core concepts and show how to bring it in action. I’ll introduce Ray tasks and actors, and show how to set up large-scale pipelines with Ray Data. As a demo project, I’ll finetune an LLM to help me with my Wheel of Time read-through. And I’ll show how can scale your workloads from a single machine to thousands of nodes. AGENDA 1. Introduction (2m) 2. Problem (3m) 3. Ray explained (10m) 4. Demo project (10m) 5. Conclusion and takeaways (5m) KEY TAKEAWAYS - For AI workloads we need frameworks that are hardware-agnostic and work on both CPUs and GPUs - With Ray tasks and actors, we can achieve distributed computing with just a few decorators - Automatic up- and downscaling is a must if you want to keep your cloud bill in check - While a literary masterpiece, The Wheel of Time has too many characters and some pacing issues PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/J3ECG7/ Planck-Bohr Rob de Wit-Liezenga PUBLISH E7QMDN@@cfp.pydata.org -E7QMDN Extending SQL Databases with Python en en 20251209T132500 20251209T135500 003000

Extending SQL Databases with Python

1. Introduction (3 min) • Who this talk is for: devs, data engineers, extension hackers • Motivation: why embed Python in databases? • Overview of the 3 systems (SQLite, DuckDB, PostgreSQL) 2. SQLite & DuckDB: Python Functions via sqlite3.create_function (7 min) • How sqlite3.create_function() works in SQLite • Example: creating a simple text-processing SQL function in Python • Use cases: rapid prototyping, lightweight data pipelines 3. PostgreSQL with PL/Python (6 min) • Enabling and using the PL/Python extension • Writing SQL functions in Python • Pros and limitations (e.g., sandboxing, permissions, virtualenvs) 4. Advanced: Embedding Python into PostgreSQL Extensions (7 min) • Writing PostgreSQL extensions in C that embed Python • Use cases: integrating ML models, custom procedural logic • Short demo or diagram: C + Python working inside Postgres 5. Trade-offs and Comparison (3 min) • Performance, deployment, complexity, ecosystem • When to use which approach 6. Q&A (4 min) • Invite questions and deeper discussion from the audience PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/E7QMDN/ Planck-Bohr Florents Tselai PUBLISH GJNEKL@@cfp.pydata.org -GJNEKL CompactifAI: Quantum-Inspired AI Model Compression en en 20251209T141000 20251209T144000 003000

CompactifAI: Quantum-Inspired AI Model Compression

We will begin with the story of how Multiverse came to be in 2019 with the mission to solve today’s problems through quantum technologies. Along this path, we discovered during a project for Bosch that quantum-inspired algorithms running entirely on classical hardware could ultra-compress AI models. In 2024, we realized that these same techniques could be applied to Large Language Models. This insight gave birth to CompactifAI. From there, we’ll walk through CompactifAI and its compression pipeline, highlighting how it outperforms naive pruning or quantization approaches in both precision and control leveraging Tensor Networks. Attendees will see how this enables new deployment scenarios: running powerful LLMs on edge devices, routing queries between local and cloud models, and even removing or restoring specific behaviors (e.g. safety filters or domain knowledge). Additionally, we’ll show how to integrate our compactifAI compressed models via API with minimal code changes and provide relevant developer resources for those interested in benefitting from faster, cheaper, more efficient models. This talk is aimed at ML engineers, researchers, and technical leads working with LLMs, vision models, or constrained deployment targets who are ready to think beyond just “bigger is better.” PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/GJNEKL/ Planck-Bohr Jon Leiñena Otamendi PUBLISH FVGMBL@@cfp.pydata.org -FVGMBL From Experiment to Enterprise: Architecting AI for Stability and Scale en en 20251209T144500 20251209T151500 003000

From Experiment to Enterprise: Architecting AI for Stability and Scale

Using the Dell AI Factory as an example, we’ll show how an open ecosystem, validated designs, and expert services help teams build, deploy, and scale AI securely and repeatably. We’ll illustrate how silicon, model, and software diversity matters and can be included in a production environment. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/FVGMBL/ Planck-Bohr Roland Kunz PUBLISH 7FTJYA@@cfp.pydata.org -7FTJYA Ingestify: Rethinking Ingestion for Complex Data en en 20251209T152000 20251209T155000 003000

Ingestify: Rethinking Ingestion for Complex Data

Ingestify is a Metadata-First Ingestion Layer that stores raw data as-is, enriched with structured metadata, and ingests only when the source data has actually changed. This makes the ingest step fast, durable and independent of downstream processing. Transformations can evolve, fail or be retried freely while the original data remains intact. In this talk, I’ll show how metadata-first ingestion forms a clean and reliable foundation for working with large, complex or non-tabular data - and why it’s a good fit for modern data workflows. PUBLIC CONFIRMED Talk https://cfp.pydata.org/pydata-eindhoven-2025/talk/7FTJYA/ Planck-Bohr Koen Vossen