Welcome And Intoduction

• Constantine Dovrolis (The Cyprus Institute)

Vectoria Your Private AI

• Luca Babetto (CINECA)

In today’s context, where companies are increasingly pushed to integrate Artificial Intelligence solutions into their processes, secure and responsible data management is a non-negotiable requirement. Many generative tools currently available operate on external infrastructures, exposing sensitive information to remote servers and creating dependencies on closed or paid models. For many organizations – from SMEs to public institutions and large industrial groups – this way of adopting AI is neither sustainable from a security perspective nor from a governance standpoint. Designed by EuroCC Italy, Vectoria was created to address the need for security in AI systems. It is designed to operate entirely within the organization’s infrastructure, without entrusting data to third parties or public cloud systems. Based on the Retrieval-Augmented Generation (RAG) approach, Vectoria enables the creation of advanced chatbots capable of understanding and querying corporate documentation in a contextual, updatable, and verifiable way – all while maintaining full control over data, models, and metrics.

Fine-tuning Large Language Models for Satellite Communications Knowledge Management: Challenges and Impacts

• Ioannis Christou (American College of Greece)

The application of large language models (LLMs) to specialized fields, such as Satellite Communications (SatCom), presents unique challenges due to the extensive and cutting-edge knowledge required. We present a fine-tuning approach for adapting 7-billion-parameter instructed LLMs (Llama-3v and Mistral) to SatCom, using a proprietary corpus sourced from the European Space Agency (ESA) consisting of domain-specific PDF documents. The confidential nature of this corpus imposes constraints on both model training and evaluation, demanding a sensible text extraction pipeline capable of handling complex structures, such as tables, to preserve critical information. Our fine-tuning methodology employs a carefully configured process, followed by an automatic evaluation framework using a curated Q&A set tailored to SatCom. Models were created in both non-quantified and 8-bit quantized formats, ensuring feasibility for desktop-level inference. The fine-tuned models demonstrated a 6,6% improvement over the baseline LLM, as well as significant gains when compared to retrieval-augmented generation (RAG) methods.

AI Agents and LLMs for Modern Cyber Defense

• Jani Dugonik (University of Maribor)

Recent advances in artificial intelligence have transformed the landscape of cybersecurity research, particularly through intelligent agent-based simulations and large language models (LLMs). Traditional cybersecurity datasets are often static and outdated, limiting the capacity to train adaptive AI systems. To address this, we developed a dynamic synthetic simulation framework that transforms static research datasets into interactive, agent-based environments. Building upon this foundation, our latest work integrates LLM-powered agents capable of generating, understanding, and classifying realistic phishing attacks in a controlled simulation. This presentation introduces the combined framework, where attacker agents simulate social engineering campaigns, and defender agents, which are supported by either classical machine learning or LLMs, analyze and detect phishing content. The integration of LLMs into the simulation framework opens new directions for adaptive cyber defense, human-in-the-loop training, and proactive threat intelligence.

LLM Steerable Computer Vision Pipeline

• Janez Perš (University of Ljubljana)

Foundation models plus ample compute make many “moderate” vision tasks solvable with minimal custom code. This talk introduces an LLM-steerable pipeline that compiles a brief YAML spec into end-to-end segmentation, zero-shot classification, and optional geometry checks, executed on GPU clusters. A remote multimodal LLM (e.g., ChatGPT) generates the configuration based on sample images and human description of the task; a Python runner on HPC invokes SAM2 for mask proposals, CLIP for prompt-driven labels, and optional BLIP-3 VQA for per-crop verification. Crucially, this workflow may double as a data engine: it produces large, reasonably clean pseudo-labeled sets with little manual effort, enabling distillation into compact models that run without HPC.

Uncovering Archaeological Sites from Aerial Imagery Using Vision Transformers

• Daniel Canedo (University of Aveiro)

Archaeological site detection is entering a new era thanks to advances in remote sensing and artificial intelligence. Archaeological sites such as hillforts often have irregular and complex shapes, making them challenging to identify using conventional computer vision methods. Multimodal approaches that combine LiDAR-derived LRM images with aerial orthoimagery improve detection accuracy, but false positives remain a major problem. In this talk, we explore how extending the principles of large language models (LLMs) to vision can address these challenges. By using cross-modal attention mechanisms, these models integrate multiple data sources, enabling precise boundary detection, reduced false positives, and scalable application across diverse landscapes and site types. A key element of this workflow is a human-in-the-loop refinement process, where archaeologists review and provide feedback on model predictions. This iterative collaboration enriches the training data, improves the system’s ability to distinguish true sites from background anomalies, and enhances overall detection reliability. Results from Northwest Iberia show a 99.3% reduction in false positives after a single refinement cycle, and nationwide deployment in England demonstrates robust performance across varied site morphologies. By combining multimodal fusion, transformer-based architectures, and expert-guided refinement, this approach delivers both accuracy and interpretability. The talk will also discuss future directions, including predictive modelling to focus searches on high-potential areas, making large-scale archaeological surveys faster and more efficient.

Enhancing Optical Character Recognition with Large Language Models

• Spiros Millas (The Cyprus Institute)

This presentation explores how Large Language Models (LLMs) enhance Optical Character Recognition (OCR) pipelines through contextual text correction, document understanding, semantic labeling, and information extraction. It will also highlight real-world use cases such as automated document processing, invoice and receipt parsing, identity verification, and multilingual text recognition. By showcasing how LLMs add intelligence and context to OCR systems, this presentation illustrates how the combination of vision and language technologies is driving more accurate, efficient, and human-like understanding of text within images and documents.

Open Discussion – Q&A

• Anna Cosp Garcia (Barcelona Supercomputing Center)