MSc
Biosketch
Sona Hasratyan is a passionate Master’s student in Mathematics in Data Science at the Technical University of Munich (TUM), with four years of hands-on experience as a Machine Learning Engineer. She proudly earned her Bachelor’s degree in Applied Mathematics and Informatics with top grades from Yerevan State University in Armenia. Sona thrives on tackling complex challenges in AI and deep learning, always eager to push the boundaries of technology and make a tangible impact.
relAI Research
Zero-Shot Object Detection in Domain-Specific Visual Environments using Foundational Vision Models
This research investigates the efficacy of foundational VLMs for zero-shot object detection in domain-specific visual environments. While large-scale multimodal models demonstrate remarkable generalization, their performance often degrades in specialized industrial contexts characterized by high information density and expert-level symbology.
Using SOTA models like Gemini 3.1 Pro, this study establishes a performance benchmark across varying levels of technical complexity. The research identifies critical bottlenecks in zero-shot spatial reasoning and investigates systematic optimization strategies, such as noise reduction and multi-channel analysis, to bridge the gap toward industrial reliability. By analyzing where general-purpose foundational models fail in expert domains, this work contributes to the development of robust, domain-adapted vision systems capable of meeting the rigorous accuracy and safety standards required for professional deployment in technical fields.