Architecting Europe-Scale E-Commerce: Bridging Hyperscale Infrastructure and Agentic Workflows

Abstract

Operating a platform for 62 million active customers across 29 markets requires reconciling 100TB of daily transactional data with strict sub-100ms latency bounds. This talk explores Zalando’s transition from traditional architectures to a decentralized, domain-oriented ecosystem. We first analyze the infrastructure sustaining this scale, focusing on the Nakadi event broker—a RESTful abstraction over Kafka that enables thousands of microservices to communicate asynchronously—and the Trend Spotter system, which captures real-time "desire-driven" signals from across Europe.

Building on this high-throughput foundation, we dive into the Zalando Assistant as a case study for deploying multimodal agents in production. We will explore the shift from simple chatbots to sophisticated LLM-powered routing and address the engineering hurdles of large-scale deployment: enforcing data boundaries with Pydantic, optimizing retrieval under heavy load, and replacing subjective "vibe checks" with rigorous, Boolean-based evaluations.

Bio

Dr. Katya Mirylenka is a Senior Applied Scientist at Zalando in Zürich, where she drives research and development for Generative AI and the multimodal Zalando Assistant. Prior to joining Zalando, she spent nearly a decade at IBM Research Europe as a Staff Research Scientist. During her tenure at IBM, she led foundational research in natural language interfaces for databases. Dr. Mirylenka holds a Ph.D. in Data Management & Machine Learning from the University of Trento and is deeply embedded in the Swiss AI ecosystem, frequently supervising PhD, MSc students and serving as an industrial mentor and lecturer.