Data Scientist Analyst

Current Vision-Language Models (VLMs) demonstrate impressive capabilities on natural images but struggle with insurance-specific aerial imagery analysis. Generic models can identify broad categories like "airplane" or "vehicle" but fail at fine-grained distinctions critical for commercial insurance risk assessment-such as differentiating wood pallets from containers, identifying specific aircraft models, or detecting dust collectors and cooling towers. Traditional approaches require 1-3 months of dedicated data science effort per object type, making comprehensive portfolio analysis impractical., This project aims to break through these limitations by developing specialized VLMs capable of understanding insurance-relevant objects and spatial relationships in aerial imagery through natural language queries, enabling risk engineers to efficiently analyze entire portfolios without building detectors for every possible object type. Main missions Develop and validate fine-grained vision-language models for aerial imagery that enable natural language-based retrieval and classification of insurance-specific objects. Key deliverables include:

• Constructing fine-grained remote sensing caption datasets using VLMs, validated through human preference studies
• Implementing multimodal structured embeddings using graph-based and scene-based representations
• Developing novel retrieval strategies (Graph-to-Graph, VectorGraph-to-VectorGraph) with Graph Matching Networks
• Creating lightweight scene-based matching approaches for resource-constrained environments
• Demonstrating significant performance improvements over baseline CLIP-based methods on insurance-relevant benchmarks

Do you have experience in Research?, * Master's degree student or recent graduate in Computer Science, Data Science, or related field

• Demonstrated research experience in computer vision, natural language processing, or multimodal learning
• Prior work on remote sensing applications, vision-language models (CLIP, BLIP, etc.), or graph neural networks highly valued
• Experience with academic writing and publishing preferred

Technical skills

Required:

• Strong programming proficiency in Python and deep learning frameworks (PyTorch/TensorFlow)
• Experience with vision-language models and transformers
• Familiarity with computer vision fundamentals (object detection, segmentation, image retrieval)
• Knowledge of contrastive learning and multimodal embeddings

Preferred:

• Experience with large language models (GPT-4, Gemini) and prompt engineering
• Understanding of graph neural networks and scene graph generation
• Familiarity with remote sensing imagery and geospatial data
• Knowledge of retrieval metrics and evaluation methodologies
• Experience with SAM (Segment Anything Model) or similar segmentation tools

Soft skills / transversal skills

• Research autonomy: Ability to work independently on complex problems while maintaining regular communication
• Analytical thinking: Strong problem-solving skills with attention to fine-grained details
• Collaborative mindset: Experience working in cross-functional teams with domain experts
• Adaptability: Comfortable pivoting between different approaches (graph-based, scene-based, hybrid methods)
• Communication: Clear documentation and presentation skills for technical and non-technical stakeholders
• Pragmatism: Balance between research innovation and practical, deployable solutions for insurance applications

Die ERGO Versicherung AG ist mit ihrer weit über 100-jährigen Erfolgsgeschichte eines der führenden Versicherungsunternehmen am österreichischen Markt. Bekannt für unsere innovativen Produkte in der Lebens-, der Kranken- sowie der Schaden- und
Unfallversicherung wollen wir auch in Zukunft überdurchschnittliche Wachstumsraten
verzeichnen. Als Teil der ERGO Group und somit der Munich Re, einem der weltweit
führenden Rückversicherer und Risikoträger, verfügen wir über einen starken finanziellen Background und umfangreiche internationale Expertise.