As an AI Researcher at Shell in Amsterdam, I design and implement advanced generative AI and multi-agent system solutions to tackle complex business challenges. My work spans from exploring state-of-the-art AI methodologies to delivering scalable, production-ready systems that enhance decision-making and operational efficiency.
Prior to joining Shell, I worked as an AI Engineer at PwC Greece, where I developed virtual assistants, AI agents, and end-to-end machine learning pipelines using Azure Machine Learning and Databricks. Following MLOps best practices, I delivered AI solutions that were both scalable and reliable for diverse business needs.
I hold an MSc in Computer Science (with distinction) from TU Delft, specializing in Artificial Intelligence with published research in computer vision and end-to-end recognition systems. I also earned a joint BSc & MSc in Electrical and Computer Engineering from NTUA, focusing on AI, deep learning, and GAN-based generative modeling. In addition, I am certified as a Microsoft Azure AI Engineer Associate and Data Scientist Associate, along with other industry-recognized credentials in AI, cloud, and MLOps.
I am passionate about bridging the gap between cutting-edge AI research and practical business applications. My hands-on approach ensures that complex technical solutions are robust, scalable, and aligned with strategic objectives. I enjoy collaborating with cross-functional teams to translate real-world challenges into impactful AI-driven products.
Project: Train a DL model for synthetic data generation for model optimization
Project: Automatic Video-Game Review Summarization
Advances in Neural Information Processing Systems (NeurIPS 2026) – Main Track
In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications – Volume 4: VISAPP, pages 393–403
FlexiBOT is a cloud-neutral chatbot built with ASP.NET Core, using retrieval-augmented generation (RAG). It features a semantic cache for quick response retrieval, integrates open-source models via Ollama, and uses RabbitMQ for event-driven processes. The system employs modular workers for document handling, with MinIO for storage, Qdrant for vector searches, PostgreSQL for data management, and Redis for caching conversation history, providing an efficient and scalable solution for conversational AI.
View on GitHub →
RAGPal is implemented as a web application with a front-end interface for user interaction and a FastAPI-based back-end for handling requests and business logic. The system utilizes Azure OpenAI API resources for chat completion and embedding generation, and the Qdrant vector database to serve as the knowledge base for storing and retrieving documents.
View on GitHub →
The goal of this project was to develop a methodology that would predict the chessboard configuration in an input image of a chessboard. Contrary to the predominant approaches, that aim to solve this task through the pipeline of chessboard detection, square localization, and piece classification, we relied on the power of deep learning models to directly predict the configuration from the entire image.
View on GitHub →
The Chess Recognition Dataset (ChessReD) is a comprehensive collection of images of chess formations that were captured using various smartphone cameras. It comprises 10,800 images from 100 chess games. The dataset features a wide range of chess piece configurations, captured under different angles and lighting conditions. The dataset includes detailed annotations about the pieces' formations in chess algebraic notation, providing valuable information for chess recognition research.
View dataset →
The goal of this project was to develop a novel architecture which would be able to generate high-resolution images conditioned on a given text description. The proposed model (TeleGAN) decomposes the difficult task of high-quality image generation, into the more manageable sub-problems of low-res black-and-white image generation, colorization, and resolution enhancement, in three consecutive stages.
View on GitHub →
Feel free to reach out by email or connect with me on LinkedIn.
