Relying on black-box AI APIs often means trading data sovereignty and cost control for convenience. This session provides a practical blueprint for architecting a private, sovereign Model-as-a-Service platform using powerful open-source models. Attendees will leave with a concrete playbook to master observability, enforce security policies, and transition from an API consumer to a platform provider.

In April 2025, a US president had Microsoft block European Court of Justice accounts overnight. No negotiation, no warning, just a switch flipped from across the Atlantic. For European enterprises handling sensitive data under GDPR, FINMA, or the upcoming EU Digital Networks Act, this isn't a hypothetical risk. It's a preview. Yet 92% of European enterprise AI still runs on US infrastructure. Most "sovereign cloud" offerings are policy promises: contractual protections that can be rewritten or revoked. Architectural sovereignty is different: your hardware, your building, your data, your AI never leaving your infrastructure. In this talk, Alexander Zehetmaier shows what a truly sovereign AI stack looks like in production: OpenAI-compatible APIs running on enterprise-owned hardware, deployed today at European companies protecting journalistic sources and trade secrets, with an 83% median cost reduction versus cloud equivalents. The session culminates in a live open-source launch: Xinity's API Gateway goes from private to public on GitHub, on stage, making a European sovereign AI infrastructure stack freely available for the first time. Key takeaways: why policy-based sovereignty is theater, what architectural sovereignty looks like technically, and how to deploy it today.

The compiler changed everything. Before it, programmers worked in assembly — painstaking, error-prone, impossible to scale. The compiler introduced a contract: if it finishes, the output works. You don't read the assembly. You trust it. That contract unlocked decades of software progress. Today, AI is writing more code than ever. But that contract is gone. Agents say "done" when the code is broken. Engineers spend more time debugging AI output than shipping. What if we could bring that contract back? In this talk, we'll explore how combining opinionated execution, reactive computation, and sound static analysis can restore deterministic confidence to AI-generated code — and what it takes to make "done" mean done again.

Superintelligence is increasingly perceived as sophisticated orchestration of autonomous, collaborating AI agents. While the potential for full process automation is widely recognized, the current immaturity of agentic systems requires a professional bridge between state-of-the-art research and reliable business integration. My approach to address this gap is by running a research lab in parallel with consulting work, allowing for the direct translation of academic findings, e.g. in automated information extraction, into production-ready enterprise solutions. The ultimate goal is Business Superintelligence: a practical reality where an ecosystem of agents understands a company’s strategy, data, and services as intimately as a human coworker can do. Achieving this requires a disciplined path: investing in agentic research, building robust data infrastructure, applications and services with rich CLIs, and taking responsible risks through well planned and executed pilots. To illustrate this, I will share the specific architectural choices and hard-won lessons from our recent retail deployments, providing a realistic look at what it actually takes to get agentic systems into production.

Most enterprises haven't shipped multi-agent AI to production yet — but the ones that have are hitting a wall nobody warned them about. The bottleneck isn't intelligence; it's trust. The moment agents cross organizational boundaries, discovery, identity, shared semantics, and secure execution all break down. An analysis of 1,600+ multi-agent execution traces found that system design issues and inter-agent misalignment account for over 76% of all failures (Cemri et al., 2025). Drawing from hands-on experience building multi-agent systems in fintech, this talk presents a four-layer "trust stack" and maps each layer to real protocols at different maturity stages: Agent discovery via A2A and Agent Cards (a Linux Foundation standard with 150+ backing organizations); zero-trust federation patterns consolidating from the ACP-into-A2A merger; Verifiable Semantics to prevent silent "semantic drift" failures; and AWCP for secure workspace delegation when agents need to compute, not just chat. Attendees will leave with a clear architectural model for layering trust into multi-agent systems, an honest maturity map — what to deploy now, what to design for, and what to watch — and concrete failure patterns to avoid when crossing organizational boundaries with agents.

Over 10,000 SAP BW systems face end-of-support by 2030 — a migration challenge so complex that traditional automation cannot scale to meet it. Manual migrations take two or more years, are error-prone, and require scarce expertise. We're using agentic AI to change that.

We'll show how multi-agent systems reverse-engineer legacy platforms, construct knowledge graphs of thousands of interdependent objects, and autonomously generate complete modernized systems — including the migration tooling itself.

Our three-phase framework — Reverse, Rethink, Rebuild — is a generalizable pattern for any complex system transformation, compressing years of work into months with automated validation at every step.

Attendees will leave with practical insights on multi-agent orchestration, knowledge graphs for spec-driven-development, synthetic data strategies for safe testing at scale, and hard-won lessons about where AI excels versus where human judgment remains essential.

Grounded in production experience, not theory.

Building a generative voice assistant demo is easy. Getting it to production is hard. Scaling it to thousands of calls per day? That's where the real learning begins. This talk shares battle-tested lessons from taking a voice AI from prototype to handling thousands of calls daily - covering the unexpected challenges that no tutorial prepared us for.