SynIRgy 2026

Dr. Dávid Zibriczky

Head of Data Science and Cognitive AI, DSV

From Evaluation to Automation: Synthetic Data for RAG Assistants and Semi-Agentic Orchestration

Optimizing conversational AI systems is particularly challenging in early deployment stages, when little or no real user interaction data is available. In this keynote, we address three key design questions for RAG assistants and orchestration services with particular attention to synthetic data generation. First, how can RAG-based assistants be optimized before real conversational data becomes available? Second, is enterprise knowledge better retrieved through one general-purpose assistant or several specialized ones? Third, how should retrieval be orchestrated across assistants in an automated and scalable way? The presentation highlights some real examples during our journey as well as opens additional questions for further discussion during the session.

Bio

David currently serves as Head of Data Science and Cognitive AI at DSV, the world's leading freight forwarding logistics company. He leads global, cross-functional teams building advanced analytics and AI solutions that drive operational excellence and innovation at scale. He holds an M.Sc. in Computer Science and a Ph.D. in Business and Management and brings over 16 years of experience in Data Science and AI, specialized in Recommender Systems and Generative AI. His work spans logistics, travel, healthcare, finance, and tech across both startup and enterprise environments. As a passionate researcher, David has co-authored more than 20 peer-reviewed publications, holds a patent in RecSys, and regularly speaks at international conferences on applied AI and machine learning.

1. A Four-Dimensional Framework for Evaluating Synthetic Data Generation Methods: Balancing Privacy and UtilityPoster
2. AMBER: A Modular Evaluation Pipeline for Business-Oriented Recommender SystemsShort Paper, Presentation
3. A Reinforcement Learning Approach to User Group Behavior ImitationPresentation
4. Flattening Realities: Why Conventional Simulation Approaches Risk Misrepresenting Children's Information AccessOpinion, Poster
5. Synthetic Compliance for Regulatory RAG: A Progressive Benchmark Suite from Simple to Complex QueriesShort Paper, Poster
6. Evaluating Multi-Hop Reasoning in RAG Systems: A Comparison of LLM-Based Retriever Evaluation StrategiesFull Paper, Presentation
7. Are Synthetic Users Rational? A Psychometric Analysis of LLM-Generated Customer PreferencesFull Paper, Presentation
8. SynCED-EnDe 2025: A Synthetic and Curated English–German Dataset for Critical Error Detection in Machine TranslationResource / Full Paper, Presentation
9. Synthetic Data for Virtual Try-On: Methodology and Lessons LearnedFull Paper, Presentation
10. Is Synthetic Data Still Fit for Purpose in the Age of GenAI?Opinion, Poster

To access the current versions of the papers, please request access on Google Drive.