Abstract. Artificial intelligence (AI) is the key to natural interaction in extended reality (XR) — but the path from an impressive prototype to a reliable production system is challenging. This talk explains why XR systems are especially sensitive to latency, VRAM, inference times, and operational costs — and why small models may be ideal for edge devices but often fail at seemingly “simple” language. Using practical examples, we explore how targeted fine-tuning can make open-source models outperform large cloud LLMs and enable truly intuitive speech interaction in XR. The talk weaves together technical metrics, real-world experience, and concrete recommendations into a coherent story: how to move AI in XR from a demo to a scalable, immersive, and production-ready user experience.

Biography. Dr. Daryoush Vaziri  is Research Group Leader for AI-based Systems at Bonn-Rhein-Sieg University of Applied Sciences in Germany. His work focuses on generative AI, domain-specific language models, and the strategic and operational integration of AI into organizational processes.
As Co-Founder of VAGO Solutions, a German AI startup focused on practical and scientifically grounded applications of generative AI, Dr. Vaziri contributes to the development of fine-tuned open-source language models for specialized domains. This includes the SauerkrautLM family, an open-source German LLM series designed for practical deployment, including in small and medium-sized enterprises.
His academic research includes projects such as ARGUS, a remote-assistance platform to support blind and visually impaired customers during in-store shopping, and Sensorium+, which explores AI-driven approaches to improve accessibility and participation in cultural environments. His broader work spans HCI and inclusive design, with an emphasis on AI-based solutions that deliver tangible societal and economic impact.