MICES 2025

Opening and welcome by the MICES organizers René Kriegler and Sebastian Russ.

The Digitec Galaxus Vector Search Journey

Abel Camacho Guardian & Joel Widmer (Digitec Galaxus)

In our effort to enhance the search experience at Digital Galaxus, we conducted dozens of experiments over recent months to introduce vector search, refine its performance, and explore its limitations. Our work highlights the importance of fine-tuned embedding models leveraging both product and behavioral data, as well as the impact of latency optimization through techniques like vector quantization and the transition from brute-force search to HNSW indexing. Beyond infrastructure, we uncovered key behavioral insights—such as when displaying zero results is preferable to low-relevance vector matches, and scenarios where traditional keyword search still outperforms vector approaches. These findings inform our ongoing journey toward building a more effective, hybrid search system.

Abel is a Senior Analytics Engineer at Digitec Galaxus with five years of experience in e-commerce, working with machine learning to improve search.

Joel is a Software Engineer at Digitec Galaxus. He holds a master’s degree in Mathematics from ETH Zurich and specializes in developing machine learning models for search.

Taking an abandoned Solr search from zero to GenAI hero 🚀

Torsten Bøgh Köster (Freelance Search & Operations Engineer)

You have a unmaintained Solr search sitting somewhere in your tech stack? You want to revamp it with shiny new GenAI features without walking through the desert of "ground work" tuning tokenization, synonyms or query reformulation? This talk got you covered!

I took a (almost) unmaintained Solr search engine from a German publisher and used off-the-shelf LLMs to tackle recall, precision, ranking and even diversity challenges. LLMs helped to improve recall and precision by pre-processing articles. By searching on extracted information only, SEO optimizations are stripped off of the articles and very precise search results are returned.

To reveal diversity in search results, LLMs cluster the search results using retrieved facets. This leads to a compact clustered search result page that highlights different aspects of a topic.This talk will guide you through the easy steps to apply GenAI with off-the-shelf LLMs to your own search system.

Torsten is a freelance engineer operating at the intersection of search, operations and observability.

Accelerating iteration speed: ranker evaluation with debiased interleaving

Leonie Brinkmann and Mika Wolle (OTTO)

In data-driven product development, traditional A/B testing often slows down iteration cycles for customer insights. In this session we introduce the innovative debiased interleaving approach, proposed by Amazon and successfully implemented at OTTO. This approach uses a within-subject design where two ranking systems are presented simultaneously to the user, allowing data collection within a few days using only a fraction of the traffic needed for traditional A/B tests.

Discover how we have revolutionized our testing methodology through technical and methodological adjustments and learn about the benefits this brings to your product development process. Be inspired by how you can optimize your own testing cycles with this approach to gain faster and more resource-efficient insights.

Leonie is Senior Digital Analyst for Ranking at OTTO Search & Navigation. She has a science background with focus on methodological research topics

Mika is Product Manager for Ranking at OTTO Search & Navigation with a background in e-commerce with foucs on Data-Science-Driven Product Development (Personalization, Recommender Systems & Search)

From Scratch to Scale: idealo's Journey in Building a Production Learning to Rank System

Gennady Shabanov and Atakan Filgöz (idealo GmbH)

This talk provides a comprehensive, real-world account of idealo's journey in successfully developing and deploying a Learning to Rank (LTR) system from the ground up to enhance product discovery for millions of users on one of Europe's leading price comparison platforms. We will chronicle the entire process, addressing key questions practitioners face: from understanding the benefits of ML-based re-ranking and how to initiate an LTR project, to navigating the complexities of offline evaluation (metrics, pitfalls, readiness for A/B testing, and challenges with click-based training data).

The session further explores achieving significant A/B test improvements, deploying and maintaining a robust system in production (including pipeline architecture examples on AWS), tackling the challenges of model retraining, and strategies for further model iteration. This is a transparent look into applied machine learning at scale, sharing practical insights and invaluable lessons learned.

Gennady has been a Machine Learning Engineer with idealo GmbH's search team for the past five years. In this role, he is responsible for developing and implementing Machine Learning solutions and the supporting infrastructure within idealo's search platform. Prior to joining idealo, he worked at Ladenzeile.de, where he contributed to a large-scale classification system for e-commerce products.

Atakan works as a Machine Learning Engineer on the search team at idealo international GmbH in Berlin. He has a few years of hands-on experience in the search domain, including a previous role at Insider, where he focused on e-commerce search. He holds a Master’s degree in Computer Science from the Technical University of Munich.

Learning-To-Rank Framework - farm your ranking models

Ilan Dubois and Marcin Gumkowski (OLX Group)

At OLX, we faced a significant challenge: the need to experiment with rankings across many countries and categories, totaling over 100. This situation became a bottleneck, hindering our Data Scientists from progressing in their work. Given our limited human resources, it was imperative to seek more automated solutions. We implemented an LTR Framework, allowing our Data Scientists to adjust configurations and select precise features and targets for optimization. The models are stored in a Model Store, where they can be deployed for A/B ranking tests.

Ilan works as Senior Machine Learning Engineer at OLX. His work focuses on supporting Data Scientists in researching, analysing, modelling, deploying and maintaining ranking models for our different use cases.

Marcin works as Senior Machine Learning Engineer at OLX. He is an engineer with experience in software development, data processing, and building ML products. Recently Marcin is focused on improving Search using ML and Cloud Computing.

The Human Element: How we Collaborate to build trustworthy Shopping Assistants

Maria Lungu & Paul-Louis Nech (Algolia)

In this talk, Maria Lungu (FX Engineer) and Paul-Louis Nech (Staff ML Engineer) share learnings from building together an AI-powered Shopping Guides feature for thousands of Algolia customers, bringing shopping guides to millions of end-users. The talk will focus not on the product we built, but rather on the friends we made along the way: how we learned to navigate this collaboration between UX experts and AI teams, each having different priorities and vocabularies. You’ll hear from our two distinct perspectives, often complementary, sometimes at odds, on crafting AI shopping experiences that truly empower users:

• Earning trust through transparency: keeping the human in control, via progressive disclosure of AI capabilities, and honest communication about current state and errors
• Designing intuitive AI interfaces, that help users understand what's happening behind the scenes without technical jargon
• Implementing effective content moderation that protects against harmful inputs and outputs
• Evaluating and optimizing AI-generated text for readability and appropriate complexity

This talk delivers practical, non-technical takeaways applicable regardless of your role in product development. Both technical and non-technical attendees will gain insights on nurturing an effective UX-AI collaboration within their organizations. We would love to continue this discussion after the talk either as a workshop session or informally at breaks with attendees.

Maria is an experienced software engineer with a growing passion for the human side of technology, focusing on building intuitive interfaces where AI and humans can collaborate seamlessly. With seven years of experience spanning financial services at Deutsche Bank and AI-powered search at Algolia, her work focuses on an intersection she finds exciting: crafting user experiences where machine learning meets thoughtful design. Her career is driven by the goal of leveraging AI capabilities to enhance human abilities rather than replace them.

Paul-Louis is a Machine Learning engineer with 10 years of experience crafting software for a global audience. Throughout his career, he has leveraged advanced algorithms to empower developers and users alike. From SwiftKey to Algolia, Paul-Louis has built tools that bridge the gap between state-of-the-art algorithms (which work in the lab) and end products (which succeed in the wild). This requires not only great technology, but more importantly great user-experience, to make sure you build something people want!

Asking Relevant Questions

Ángel Maldonado (Empathy Holdings)

Can AI complement relevance with interrogation? Can questions become answers?

The way we hone a question defines the fortune of having a relevant answer. In this talk, Angel will present a new commerce search pattern inspired by Socratic interrogation that leverages self-hosted and open-sourced LLMs. Three interrogative commerce search experiences running on a private cloud will be discussed:

• 2D question experiences for marketplaces
• Question led discovery on foods
• Interrogative experiences for books

Ángel is a computer scientist with a passion for philosophy and design. He is CEO & founder of EmpathyHoldings, home of:

• MotiveMarket.com: Consumer Marketplace for local shops
• VisiblePrivacy.com: Consent & Privacy Compliance
• EthicalAliance.co: Ethical & Sustainable AI
• Apisearch.io: Search Solution Mid-market
• Empathy.co: Search Platform Enterprises
• Motive.co: Search Plug-in SMEs

Personalizing E-commerce Search with Latent Behavioral Embeddings

Trey Grainger (Searchkernel)

Of all types of search, e-commerce search is usually the best situated to utilize behavioral signals (queries, clicks, purchases, etc.) to interpret query intent and personalize search results. Yet, with the meteoric rise of vector databases and semantic search approaches using content-based embeddings, query intent models integrating user signals have largely been overlooked in recent years for search and RAG systems.

This oversight is understandable, as personalized search requires additional work collecting and processing clickstream behavior versus just “picking the latest LLM” and plugging it in to encode multimodal text and images. As we think of multimodal search, however, user-behavioral embeddings provide a critical additional modality for understanding users’ personal goals and interests and thus properly interpreting their query intent.

In this talk we’ll walk through practical techniques for incorporating user context and behavioral signals to enhance e-commerce search relevance, moving beyond just traditional lexical and content-embedding methods. We’ll cover how to train latent behavioral embedding models using user signals, implementing personalized search experiences with appropriate contextual guardrails to prevent over-personalization, and enabling cross-modal personalized search combining content-based embeddings with behavior-based embeddings. We’ll walk through open-source code examples with an e-commerce dataset, demonstrating how modern hybrid search approaches can integrate these behavior-based modalities to significantly improve the relevance of e-commerce search.

Trey is the founder of Searchkernel, a software company building the next generation of AI-powered search. He is an advisor to several startups and adjunct professor of computer science at Furman University. He previously served as CTO of Presearch, a decentralized web search engine, and as chief algorithms officer and SVP of engineering at Lucidworks, an search company whose search technology powers hundreds of the world’s leading organizations. He is also the co-author of Solr in Action (Manning, 2014), the leading book on Apache Solr. Trey has over 18 years of experience in search and data science, including significant work developing semantic search, personalization, and recommendation systems, and building self-learning search platforms leveraging content and behavior-based reflected intelligence. This work resulted in the publication of dozens of research papers, journal articles, conference presentations, and books at the cutting edge of intelligent search systems.

One Mixture of Encoders retrieval vs many rerankers: A modern approach to search and recommendations

Filip Makraduli (Superlinked)