A connected research trajectory.
The work spans four connected families. Each section brings together positions, methods, software, research prototypes and collaborative projects rather than presenting them as isolated outputs.
AI in EducationPOSITION · PEDAGOGICAL VALIDATION · CLASSROOM PRACTICE
My work in AI and education begins with a clear position, develops into a framework for pedagogical responsibility and returns to the classroom through open, teacher-led resources.
01 · THE POSITION
In favour of AI in education. Under clear conditions.
I support AI wherever it helps a child understand, question and create. But pedagogical decisions, responsibility and the authority to stop must remain in human hands. The manifesto connects this position to teacher agency, child protection, transparency, accountability and institutional control.
The central question is not simply whether AI enters education. It is who frames its voice, who validates its output, who controls the infrastructure and who can explain a decision before it affects a learner.
02 · THE FRAMEWORK
EPVL Ethical Pedagogical Validation Layer
EPVL sits between the teacher, the model and the child. AI-generated material remains a draft until a teacher examines its developmental appropriateness, cultural sensitivity, semantic fidelity and ethical transparency.
The workflow makes pedagogical intent operational. It records the purpose, model, changes and approval, while preserving the teacher’s authority to correct, reject or approve an output before classroom exposure.
03 · THE CLASSROOM
AI in Primary Education From principles to classroom practice.
The open-access book brings these principles into practice through twenty teaching proposals for primary education. AI operates under the teacher’s control and supports inquiry, comparison and creation rather than replacing interpretation or judgment.
The teacher operates the system, no child’s personal data enters a prompt and every output is checked before presentation. The result becomes material for questioning and revision, and AI is withdrawn whenever it adds no genuine learning value.
Visual intelligence and image retrievalDEEP EMBEDDINGS · DESCRIPTORS · RETRIEVAL SYSTEMS
This research spans two technological eras: the early development of compact, handcrafted visual descriptors and retrieval systems, followed by today’s work on deep networks, Vision Transformers, CLIP and universal image embeddings.
PRIMARY RESEARCH HUB · TODAY AND THE EARLY YEARS
Image Retrieval Tools
The complete research page brings together the current work on learned and universal image representations with the earlier descriptors, libraries, software and experimental retrieval systems that established the foundations of this research trajectory.
LATEST FINDINGS · 2021-PRESENT
2023 · IEEE ACCESS
Universal Image Embedding
A multi-domain CLIP fine-tuning strategy that retains existing knowledge while producing a transferable image encoder for unseen retrieval and recognition domains.
2021 · IEEE DCOSS
Vision Transformers for Image Retrieval
A plug-and-play global descriptor derived from a pretrained Vision Transformer, evaluated without task-specific training across established retrieval benchmarks.
2021 · EXPERT SYSTEMS WITH APPLICATIONS
Deep Convolutional Features
A systematic study of global and local representations from five deep convolutional architectures, establishing the baseline for the later transformer-based work.
EARLY FOUNDATIONS · 2008-2019
Compact Composite Descriptors
Compact visual representations combining colour and texture information while keeping storage requirements low.
SIMPLE Descriptors
Global MPEG-7 and compact descriptors applied to salient local image regions for efficient retrieval.
img(Rummager)
A complete image-search environment supporting multiple descriptors, indexing, real-time feature extraction and retrieval evaluation.
MPEG-7 Visual Descriptors
Open-source C# implementations of the Scalable Color, Color Layout, Dominant Color and Edge Histogram descriptors.
img(Anaktisi)
A web retrieval system built around compact colour and texture representations ranging from 23 to 74 bytes per image.
MMRetrieval.net
An experimental multilingual and multimodal search engine combining independently indexed modalities through configurable fusion methods.
Autonomous systems and roboticsMULTI-ROBOT PLANNING · AERIAL · UNDERWATER · EDUCATIONAL ROBOTICS
My contribution spans a continuous line from foundational multi-robot planning algorithms to integrated aerial and underwater platforms, real-world UAV operations and educational robotics. Across this work, I have contributed as a co-author, senior researcher and technical scientific manager.
FOUNDATIONAL CONTRIBUTION · MULTI-ROBOT COVERAGE
DARP: Optimal Multi-Robot Coverage Path Planning
I co-developed DARP, an algorithm that divides a terrain into connected, robot-exclusive regions according to the robots’ initial positions. Combined with Spanning Tree Coverage, it enables complete, non-backtracking coverage, balanced use of the team and optimal coverage time whenever an optimal solution exists.
FP7-ICT ROBOTICS · SENIOR RESEARCHER
sFly: Vision-Controlled Micro Flying Robots
As a member of the sFly research team and co-author of the integrated system paper, I contributed to work combining onboard monocular SLAM, inertial sensing, 3D map merging and coordinated coverage. The project demonstrated three micro-aerial vehicles navigating autonomously and mapping an unknown GPS-denied environment.
FP7-ICT ROBOTICS · TECHNICAL SCIENTIFIC MANAGER
NOPTILUS: Autonomous Underwater Systems
As Technical Scientific Manager of NOPTILUS, I worked on autonomous coordination for teams of underwater vehicles in unknown and changing environments. This research line later produced a distributed plug-and-play algorithm in which each robot learns how its actions contribute to a team objective, without requiring an a priori analytical model and while retaining operational constraints and fault tolerance.
REAL-WORLD UAV OPERATIONS · 2023
Efficient Coverage Path Planning for Real-World UAV Missions
I co-authored a grid-based Coverage Path Planning methodology with three application-specific modes: Geo-fenced, Better Coverage and Complete Coverage. Evaluated on 20 benchmark regions, it supports complex concave areas with obstacles while enabling strict geo-fencing, complete coverage, non-overlapping trajectories and paths without sharp turns.
EDUCATIONAL ROBOTICS · ROBOTEX CYPRUS
Robotics Competitions and Learning Outcomes
I co-authored a framework that links educational robotics platforms and competitions to six expected learning outcomes, while classifying platforms by the skills they require rather than manufacturers’ age ranges. This research complements nine years of service on the scientific team of Robotex Cyprus.
Applied AI, learning and emerging systemsMACHINE LEARNING · SMART AGRICULTURE · STRUCTURAL MECHANICS · BLOCKCHAIN
This group connects computational methods with practical challenges in machine learning, smart agriculture, structural mechanics and decentralised systems.
MACHINE LEARNING
Gradient-Free Training of Artificial Neural Networks
A numerical method for computing neural-network weights without a laborious iterative training procedure, evaluated on demanding regression and classification problems.
SMART AGRICULTURE
Precision Agriculture
My work in precision agriculture connects field sensing, computer vision and low-power communications. One study combined LoRa transmission with a fine-tuned CNN and Grad-CAM to identify grape leaf diseases from low-resolution and visually corrupted images. An earlier field-tested system used an automated McPhail trap, remote imaging and automatic insect counting to monitor the olive fruit fly, achieving almost 75% counting accuracy.
HIGH-PERFORMANCE MACHINE LEARNING · STRUCTURAL MECHANICS
High-Performance Machine Learning for Structural Mechanics
I co-authored a generic parallel and distributed machine-learning framework for computationally demanding structural mechanics problems. Four algorithms were evaluated across three structural engineering applications, with the best-performing XGBoost-HYT-CV model achieving a 4.54% residual error and, in one case, reducing the error threefold against existing methods.
DECENTRALISED SYSTEMS
Research on Blockchain
Research on application-driven blockchain systems and specialised decentralised applications, including Parkchain, RandomBlocks and mechanisms for sensitive information exchange.