Coffee served.

As AI capabilities continue to advance globally, digital sovereignty is becoming increasingly important for organisations and societies. In this talk, Prof. Lokke Moerel will explore the latest AI developments and the challenges they create for Europe’s digital independence. She will highlight EU initiatives designed to strengthen digital sovereignty, how organisations can benefit from these efforts, and the current progress of GPT-NL — the first sovereign Dutch LLM, developed to align with European and Dutch legislation and public values. The session will also explore the opportunities of using GPT-NL for agentic AI applications and conclude with practical ways organisations and individuals can contribute to a more sovereign digital future.

Speaker: Prof. Lokke Moerel - Professor Global ICT Law at Tilburg University
Moderator: Leanne Joseph

The next wave of AI innovation will not be defined by prototypes, but by the systems that successfully make it into production. In this keynote, Pratyush explores what it takes to turn AI agents from promising demos into reliable, scalable products. Drawing on hands-on experience building production-grade agentic systems, he will share the architectural principles, engineering practices, and hard-earned lessons behind deploying AI at scale.

From LLMs acting as application compilers to GitOps-driven workflows and robust delivery architectures, Pratyush will reveal how organisations can design AI systems that are faster to iterate, easier to maintain, and ready for real-world use. The keynote highlights why production discipline - from clear contracts and safeguards to thoughtful system design is the key ingredient that transforms AI capabilities into meaningful products.

Speaker: Pratyush Sinha - Technical AI Executive, Decathlon
Moderator: Leanne Joseph

A dedicated moment to connect with fellow AI professionals, exchange ideas, and build new relationships. During this interactive networking activity, participants will meet others from the community, share experiences, and discover opportunities for collaboration across the AI ecosystem.

Agents look great in demos and fall apart in production, and most teams have no reliable way to tell the difference. In this hands-on workshop, Orq.ai walks through simulating, evaluating, and debugging agent behavior at scale, using the same tooling used to ship agents into real production systems.

Facilitator: Bauke Brenninkmeijer - AI Research Engineer at orq. ai & Arian Pasquali, Reserach Engineer at orq.ai

Reliability is the bedrock of modern software, and providing it at planet-scale is a great engineering challenge. In this talk we'll discuss Site Reliability Engineering (SRE): what it is, what we do, and how we define reliability. We'll take a journey through the evolution of Google Search, and the lessons learned from keeping its planet-scale infrastructure stable around the clock. Finally, we will explore the emerging frontier of AI for SRE, looking at how tooling and automation can reduce operational toil and shape the next generation of resilient systems.

Speaker: Almer S. Tigelaar - Staff Engineering Manager & Tech Lead, Google
Moderator: Leanne Joseph

This talk presents an integrated framework combining Specification-Driven Development (SDD) and Test-Driven Development (TDD) for AI-assisted (GitHub Copilot) software development within Agile environments. An integrated methodology benefits from the strengths of both approaches. Specifications serve as both human documentation and structured directives for AI agents. Well-defined tests act as validators of requirements and catch potential defects in an early stage.

Agile ceremonies provide natural integration points for an integrated approach: Refinement sessions can be used to collaboratively develop specifications and test cases, for which we have defined a new Definition of Ready (DoR). Sprint reviews can be used to validate that the implementation meets the specifications and that known edge cases are covered in tests.

Speakers: Liangliang Zheng - Data/MLOps Engineer, ING & Kevin Kewal - Engineering Lead, ING

Most teams building LLM-powered data agents start from scratch — generating chart code, building new visualisation layers, replacing existing dashboards entirely. We took the opposite approach, and it worked.

This talk presents the AI assistant we built for SensingClues, a conservation technology NGO, to help field practitioners interpret satellite vegetation and fire data without needing to navigate a complex dashboard manually. The assistant is embedded in an existing R Shiny application and powered by a tool-calling agent over a FastAPI service — built entirely by volunteers through CorrelAid Netherlands.

The core architectural insight: instead of having the LLM generate charts, we had it return structured references that trigger the app's existing Plotly, Leaflet, and static image functions. The dashboard became the rendering engine. The agent became the reasoning layer. No chart code generation. No visualisation regression. If your charts are defined as code functions - in Shiny, Streamlit, Dash, or any custom web application — they can become agent-callable without a rewrite.

Speaker: Felicity Fan - Data Scientist, CorrelAid Netherlands

Environmental crime is often prosecuted through a side door statute, the "Al Capone" approach. Finding those angles means knowing which laws, rulings and EU statutes could apply, including the less obvious ones, and that knowledge sits scattered across separate corpora and in the heads of a few experienced investigators. KickstartAI team built a retrieval engine for the Dutch National Police that turns an intake document (preweegdocument) into ranked legal angles. The catch: a system whose value is surprise breaks a good part of the standard retrieval playbook.

What the team built: an uploaded pre-assessment document goes through two LLM passes. The first extracts a structured case frame, the second generates queries along four routes: explicit references, direct legal grounds, lateral angles, and creative "Al Capone" angles. Hybrid search (BM25 plus dense vectors, RRF) then runs across five per-source indexes covering 3.2 million passages, 69 GB of Dutch and EU legal text. Results are deduplicated, aggregated by law, and streamed into a step-by-step wizard. A key design choice is that results from different search routes are never globally reranked, preserving the surprising and unconventional findings that make the system valuable.

The talk will explain the reasoning behind this approach, share findings from 382 blind relevance judgments by six detectives, and demonstrate the system live.

Speakers: Henk Snepvangers - AI Project Lead, KickstartAI & Sanne Wielinga - Senior ML Engineer, KickstartAI

Everyone's experimenting with AI. Far fewer are running it in production. Closing that gap is one of the biggest challenges organisations face today, and it's exactly what this hands-on workshop is about. Participants will design, build, and deploy enterprise AI agents on the Wonderful Platform.

Starting from a real enterprise use case, you'll follow the full journey from an initial idea to a fully operational AI agent. Along the way, you'll tackle what actually separates a demo from a deployable system: system integrations, governance, testing, observability, and making agents reliable and trustworthy in practice.

Through practical exercises and a live demonstration, you'll discover how to build agents that go beyond prototypes and turn AI into something organisations can genuinely put to work.

Hosted by wonderful.ai
Facilitator: Sako Arts - CTO Wonderful NL

Building a production image generation pipeline sounds straightforward, until real-world complexity hits. In this session, we share hard-won lessons from designing and operating an end-to-end image generation pipeline at IKEA, processing over one million images per month.

We'll walk through the full journey: handling noisy, real-world customer images, building robust evaluation frameworks, fine-tuning our own models, integrating foundational models, and scaling serving infrastructure on the cloud.

What worked, what failed, and what we'd do differently if we would have started all over.

Speaker: Shahin Shahkarami - Global VP of AI & Automation, IKEA Retail
Moderator: Leanne Joseph

As AI moves from prototype to production, many teams discover that cloud-only architectures do not fit their security, latency, compliance, or data-sovereignty requirements. In this session, Canny shares practical lessons from building and operating an AI stack that can run on-premises: from model serving and RAG pipelines to vector databases, observability, evaluation, secrets management, GPU constraints, and upgrade strategies. Rather than treating on-prem as a special-case deployment, we will look at how to design AI systems that are portable, secure, and maintainable across customer environments. We will discuss trade-offs between managed services and self-hosted components, how to keep sensitive data under control, and what it takes to support production AI in regulated or sovereign environments.

Sebastiaan van Duijn - AI Manager, Trivium Packaging

Brainbay's AI agent AskNVM came to life in early 2025. At that time, it was still hard to get agents to follow basic tasks. Hence, they engineered an agent that followed a graph in order to stay as close to our instructions as possible. As models got better, however, the world moved towards ReAct (reason + act) agents, and so did they. Overnight their agent went from being a hallucination specialist to a helpful and powerful assistant to 10.000 real estate brokers in the Netherlands. In this session, Tom will discuss how to build agents in production. He will address the impact of the ReAct agent loop and what it means for you as a builder. Also, he will discuss the role that traditional graph-based approaches still have. You will learn how to use an engineering mindset to determine which features/hypes (multi-agents, reasoning, MCPs, skills and the like) your agent needs. Spoiler alert: you are not smarter than modern LLMs, so let them cook!

Speaker: Tom Kersten - LLM Engineer, Brainbay

Frontier LLMs keep getting better and cheaper: GPT-5 nano costs $0.05 per million input tokens and Gemini 2.5 Flash Lite is $0.10. At these prices, is there still a case for running your own small models?

If you grab an off-the-shelf small model and point it at a production task, the answer is no. Base small models are simply not good enough. But there’s a third option that most teams overlook: fine-tuning changes everything. A small model that has been fine-tuned on your specific task doesn’t just close the gap with frontier LLMs, it matches or beats them, while running 10x cheaper on your own hardware.

We tested this across 8 datasets, comparing fine-tuned small models (0.6B to 8B parameters) against 10 frontier LLMs from OpenAI, Anthropic, Google, and xAI. The fine-tuned models ranked first overall on half the tasks and placed 3.2 on average, just behind Claude Opus 4.6 (2.5), while beating out Gemini 2.5 Flash (3.5) at 100x lower cost.

Speaker: Usman Zafar - Machine Learning Engineer, Distil Labs

Running LLMs on consumer hardware sounds simple download a model, run llama.cpp, done. The reality is messier, and more interesting.

This talk has two interleaved threads. The first is the math: how quantization actually works mapping float32 weights to int4/int8 via scale factors and zero points, what the rounding error looks like in practice, and why the compute graph changes shape during inference in ways that affect latency beyond just memory savings. The second thread is grounded in shipping TweeTaal, a production iOS app that runs a fine-tuned multilingual model entirely on-device via MLX no cloud, no API calls, no user data leaving the phone.

Together they answer the questions that only arise when theory meets deployment: which quantization level preserves task quality without breaking things quietly, why memory bandwidth dominates inference speed more than raw compute and where Apple Silicon's unified memory architecture changes the story entirely.

Speaker: Sarthak Anand - Senior Data Scientist, Libra AI

Moderator: Leanne Joseph

You may already have read about it on NOS.nl or in the newspapers over the past few weeks: an AI tool has been developed to predict which players are at risk of developing gambling problems. The tool was published by the Dutch Gambling Authority on 18 August and was discussed in Parliament on 3 October.

The creator of the tool will share:

• His personal motivation and experience with gambling addiction.

• How a mysterious email about a largely forgotten law ultimately led him to leave his job at bol.

• The challenges of working on a project that could have significant financial consequences for the gambling industry.

• How AI can be used to create genuine value, and how we can protect ourselves from the marketing hype surrounding “AI”.

Speaker: Charles de Leau - Data Scientist, UvA

Everyone has an AI strategy. Far fewer can say what actually works once it meets real teams, real data and real governance. In this fishbowl, AI leaders share how they apply AI inside large organisations: what delivers value, where they get stuck, and what they would do differently. There's no stage, no slides and no panel. Just a few open chairs in the middle of the room: take one to share what works, what doesn't, or to ask the room a question.

Moderated by Nic van Vliet and Bartjan van Langeveld, KickstartAI

How can AI contribute to solving one of the world’s biggest challenges: ensuring access to safe, nutritious, and sustainable food for all?

In this interactive session, Zero Hunger Lab explores how data and AI can support progress towards SDG 2: Zero Hunger. The session starts with a look at the current state of global food security and introduces the key challenges behind achieving zero hunger - from ending malnutrition and improving agricultural productivity to building climate-resilient food systems and protecting biodiversity.

Through interactive polls and real-world examples, participants will discover how AI is already being applied to address these challenges and explore where technology can make the greatest impact. The session concludes with a discussion on how individuals and organisations can contribute to creating a more sustainable and equitable food system.

Speaker: Frans Cruijssen - Zero Hunger Lab, Tilburg University

After a full day of watching AI's potential ends with Marco van der Werf asking what it will lead to. He paints two visions of the Netherlands in 2041. In one, care is more affordable, queues are shorter, and we build rather than rent intelligence. In the other, none of that happens. Not because someone stops it, but because no one steers it. He then brings it back to this room and the Monday morning decisions that could shape which future we move toward.

Speaker: Marco van der Werf - Founder at LenAI and Fellow at the UvA
Moderator: Leanne Joseph

Wrap up the day by connecting with fellow AI professionals over drinks and bites. This informal networking moment offers the opportunity to continue conversations, exchange ideas, and build meaningful connections with members of the KickstartAI community.