Talk and breakout sessions 2026
Panel Discussion | LLMs in the Wild: AI at Scale Across Enterprise, Industry & Open Source
Room: Grote zaal – 2nd floor
Time: 11:15 – 11:45
Moderator
Jessica Greene, Senior Machine Learning Engineer, Ecosia GmbH
Speakers
An Dan, AI Engineer, Lufthansa Industry Solutions
Tereza Iofciu, AI & Leadership Coach,
Nico Martin, Open Source Machine Learning Engineer, Hugging Face
Daniel Zeitler, Creative Technologist
Unchecked traffic congestion can render our microservices unresponsive, driving away customers – we do not want our users to ever see an HTTP 503. So, how do we (a) detect congestion before it breaks end-user experience and (b) recover from congestion? To make congestion detection and mitigation resilient and scalable, we must also be able to do it with local measurements inside our application.
Microservices are sandwiched between two pieces of the congestion problem: (a) users may start making too many requests and (b) services we depend on may become slow. Either of these scenarios can cause a traffic congestion. We want to detect this and gracefully degrade our quality of service to allow users to know what is going on, instead of an unexplained HTTP 503 status.
In this talk, I will explain how we can build resilient micro services with built-in congestion control mechanisms. I will use a sample Java Spring Boot application to demonstrate the effect of congestion building up and illustrate how we can detect and react to it. All of this will be done with simple local measurements to ensure stability and resilience.
The Sovereign Stack: Building Beyond Hyperscalers in the European Cloud
Room: Grote zaal – 2nd floor
Time: 13:30 – 14:00
Adriaan Wind, CTO & Co-Founder, anny
We all know the giants: the US-based hyperscalers that dominate the SaaS, PaaS, and IaaS markets. Developers love their DX and endless feature lists, but the mindset has shifted. It’s no longer just about checking a GDPR box; it’s about digital sovereignty. As dependencies grow and geopolitical landscapes shift, relying solely on non-European infrastructure has become a strategic risk—especially for those working with the public sector or critical infrastructure.
But is it actually possible to build a modern, scalable SaaS solution on a purely European foundation? One where not just the data center, but the parent company and the entire jurisdiction remain within the EU?
At anny, we decided to find out. We launched an ambitious experiment: building a parallel environment from the ground up using exclusively European providers. In this session, we’ll share our journey—the technical hurdles, the unexpected wins, and the reality of achieving true strategic autonomy.
Your Integration Layer Is Already Legacy - You Just Don't Know It Yet
Room: IJ zaal – 5th floor
Time: 13:30 – 14:00
Alex Tryshchenko, Senior Engineering Manager, Typeform
Most integration ecosystems are built for one job: a user connects a tool, a webhook fires, data moves downstream. That contract holds until the product need appears for something fundamentally different from the same plumbing.
At Typeform, we build and maintain 45 native integrations alongside a long tail of third-party providers we don’t control. Over the past two years we’ve had to rearchitect how that ecosystem works under compounding pressure: powering event-driven automations not just form submissions; supporting inbound data flows alongside outbound webhooks; routing around EU data residency constraints on connectors we don’t build; governing a surface area that grows faster than the team that owns it; and making integrations safely callable by AI agents without a human approving each step.
This talk covers what it actually takes to evolve an integration layer under that kind of load. We’ll walk through the architectural differences between inbound and outbound integrations and why conflating them causes production failures. How we approached EUDC compliance across third-party connectors, including the routing decisions that bought us time. How we govern an ecosystem that grows faster than any single team can review. And where AI agent invocation fits in as one pressure among several, not a rewrite trigger on its own.
If your integration layer is holding together today but you’re not sure it’ll survive the next two years, this talk is for you.
How to bootstrap a fully managed CAPI environment using gitops
Room: Grote zaal – 2nd floor
Time: 14:05 – 14:35
Farnaz Babaeian, Senior Platform Engineer, Leaseweb
João Falcão, Senior Platform Engineer, Leaseweb
Sohan Maheshwar, Lead Developer Advocate, AuthZed
Room: Grote zaal – 2nd floor
Time: 13:50 – 14:20
Managing Kubernetes clusters at scale is hard. Doing it consistently, repeatably, and without snowflakes is harder. Cluster API (CAPI) promises declarative, Kubernetes-native cluster lifecycle management; but standing up a production-ready CAPI environment still involves a lot of moving parts: bootstrap providers, infrastructure providers, management clusters, and the chicken-and-egg problem of who manages the manager.
In this talk, we’ll walk through how to bootstrap a fully managed CAPI environment from the ground up using the CAPI Operator and ArgoCD. By adopting the Operator, we replace imperative clusterctl workflows with a fully declarative model for the CAPI lifecycle. This allows your entire ecosystem – from the management cluster and CAPI components to the workload clusters themselves, to be defined in Git and continuously reconciled by ArgoCD.
We’ll cover:
– The CAPI architecture and why a GitOps-driven approach removes operational toil
– Bootstrapping the management cluster using tools like kind and handing it off to ArgoCD
– Installing and managing CAPI components declaratively using the CAPI Operator, keeping provider lifecycle fully under GitOps control
– Handling the pivot: moving from a temporary bootstrap cluster to a self-managed management cluster
By the end of this session, you’ll have a clear mental model and a practical blueprint for running a CAPI environment where Git is the single source of truth, human toil is minimised, and cluster provisioning becomes a pull request.
Whether you’re running on bare metal or a public cloud, the patterns shown here apply across infrastructure providers.
Why is this slow, expensive, and wrong? Facilitating resilience through observability
Room: IJ zaal – 5th floor
Time: 14:05 – 14:35
Felix Meyer, Chapter Lead – Operations Engineering, Cloudeteer
AI workloads fail in ways traditional monitoring often misses. Kubernetes can report healthy pods while GPUs sit at 98% utilization. HTTP 200 responses can hide seven-second queue waits. Inference costs can double overnight without a pod restart, failed request, or obvious SLO violation. The system is “up” — but slow, expensive, or wrong.
This talk presents a practical observability architecture for resilient AI inference platforms. We will trace one request across five layers: API Gateway, LLM Proxy, Message Queue, Inference Server, and LLM Observability platform. Using OpenTelemetry as the instrumentation backbone, we will show how a shared trace_id connects infrastructure telemetry, inference metrics, and LLM-specific signals such as queue_wait_ms, time_to_first_token, tokens_per_second, gpu_utilization, model_name, cost_per_request, and quality_score.
The focus is operational: how do platform teams detect failure before users notice? We will walk through concrete feedback loops such as scaling inference capacity when queue latency rises, triggering model fallback when GPU saturation causes degraded latency, and alerting when quality scores regress while HTTP responses remain successful. The goal is not another dashboard, but telemetry that can drive automated infrastructure decisions.
The session includes a live demo using Grafana, Tempo, and Langfuse. We will simulate common AI platform failure modes and use traces, metrics, and LLM observability data to answer three production questions: Why is this slow? Why is this expensive? Why is this wrong?
Attendees will leave with a practical mental model for building AI infrastructure that is observable, resilient, and scalable under real load.
The Invisible Infrastructure: How AI Systems Eliminate Engineer Choices Before Anyone Decides
Room: Grote zaal – 2nd floor
Time: 15:15 – 15:45
Jacob Ward, Founder and Journalist, TheRipCurrent.com, CNN
Every engineer working with AI-powered infrastructure today is making decisions inside a system that has already made most of the decisions for them. Not through malice — through design. The models you integrate, the APIs you call, the deployment architectures your vendors recommend: each one narrows the option space before your team ever convenes a meeting. This is what I call The Loop — and after twenty years covering AI development from inside the labs, boardrooms, and newsrooms where it was built, I can tell you it’s not a side effect. It’s the product.
This talk is a technical map.
I’ll walk through the specific mechanisms by which AI systems reduce the surface area of engineering decisions — in infrastructure resilience, in automation tooling, in vendor lock-in that masquerades as capability. I’ll show how the same cognitive vulnerabilities that AI systems are now designed to exploit in consumers also operate on engineering teams under deadline pressure. And I’ll lay out what resilient AI-integrated architecture actually looks like when you design against these dynamics rather than into them.
The engineers who build the next generation of infrastructure will either inherit systems that narrow their choices or design systems that preserve them. This talk gives you the framework to tell the difference — and the technical instincts to build toward the latter.
Takeaway: a practical decision-audit framework for AI integrations that flags where your architecture has already ceded control, and where it doesn’t have to.
Oops!… We F*ed It Up Again: The Art of Staying Calm When Systems Crash
Room: IJ zaal – 5th floor
Time: 15:15 – 15:45
Alon Kiriati, Tech Lead & Principal Engineer, Autodesk
No matter how robust your code or how comprehensive your tests, systems will fail—it’s a certainty of software engineering. The key to resilience is not avoiding failure but ensuring rapid detection, effective response, and comprehensive recovery. In this talk, we’ll explore the core principles of on-call rotations and their critical role in handling incidents.
We’ll dive into the lifecycle of an emergency—from detection to restoration—covering essential practices like monitoring, alerting, and ownership. Learn why on-call systems are vital, the dangers of over-relying on senior engineers, and how to empower teams with actionable playbooks, feature flags, rollback strategies, and data restoration techniques. Whether you’re designing your first on-call rotation or optimizing an existing one, this talk provides practical steps to transform chaos into confidence and ensure your business is resilient in the face of failure.