CV

Research Intern

TNO - Netherlands Organisation for Applied Scientific Research

Worked on improving In-Context Learning in Computer Vision with Patch Nearest Neighbor Consistency as part of my Master's Thesis.

• Collaborated with Dr. Yuki M. Asano and Dr. Gertjan Burghouts.
• Advanced understanding of in-context learning for vision tasks.

Data Science Intern

South Pole

Enhanced product-to-category matching algorithms using machine learning techniques.

• Diagnosed issues in the matching of products to CO2 footprint algorithms through log analysis.
• Created a ground truth dataset for evaluating and training a better matching of Products to Categories via web scraping.
• Improved top-1 system's Product to Category matching accuracy by 25.2% using a Text Transformer Encoder.
• Deployed models to production with Kubernetes.

Research Intern

CYENS Centre of Excellence

Developed a Python library to accelerate Quality Diversity Optimization algorithms.

• Enhanced efficiency of state-of-the-art QD algorithms by over 2x using JAX.

Software Development Engineer Intern

Amazon Data Services Ireland

Developed tools to optimize internal data processes and improve code logging practices.

• Automated report generation, saving hours of manual effort.
• Improved a Java library for accessing internal AWS data.
• Promoted good code logging practices, saving thousands of dollars annually.
• Optimized a large-scale real-time log analysis service.

Research Intern

CYENS Centre of Excellence

Built a Speech Emotion Recognition system for dyadic conversations.

• Engineered an end-to-end system using state-of-the-art methodologies.

Co-Founder & Backend Developer

Fooderloo

Co-founded a business addressing food waste in Cyprus, developing its backend systems.

• Designed and implemented backend, database, and server-side systems.

Research Intern

Department of Computer Science, University of Cyprus

Developed algorithms to improve micro-network performance and fault tolerance.

• Designed fault-tolerant message routing and faulty node detection algorithms.
• Developed a Java-based simulation program for testing routing algorithms.

Signal Soldier

National Guard of Cyprus

Served as a Network and Computer Technician, operating intra-office communications.

PhD

University of Technology Nuremberg (UTN)

Deep Learning for Vision

MSc

University of Amsterdam (UvA)

Artificial Intelligence (AI)

• Deep Learning 1 & 2
• Computer Vision 1 & 2
• Natural Language Processing
• Machine Learning
• Game Theory
• Information Retrieval
• Fairness, Accountability, Confidentiality and Transparency in AI
• Knowledge Representation and Reasoning

BSc

University of Cyprus (UCY)

Computer Science (CS), with a focus in AI

• Data Structures & Algorithms
• Object Oriented Programming (Java)
• Algorithms
• Computer Networks
• Operating Systems
• Databases
• Software Engineering
• Artificial Intelligence
• Machine Learning
• Computer Vision
• Data Mining on the Web
• Calculus I & II
• Linear Algebra
• Probabilities & Statistics
• Theory of Computation
• Introduction to Economics
• Logic Programming & Artificial Intelligence
• Parallel Processing
• System Security
• Synthesis of Parallel Algorithms
• Programming & Tools
• Logic in Computer Science

Best Ready for Production Solution

Summer of AI, initated by ABN AMRO and DEUS

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

1 st place in the "Go Green, Go Digital, Go Start-up!" competition

University of Nicosia

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

1st place in the Cyprus Digital Championship Competition for Student Entrepreneurship

Digital Championship Committee (Ministry of Energy, Commerce and Industry; Ministry of Education, Culture, Sport and Youth; Institution of research and innovation

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

Most Innovative idea for the society awarded with 'Innovative Technologies for the Society' in the Cyprus Digital Championship Competition for Student Entrepreneurship

Digital Championship Committee (Ministry of Energy, Commerce and Industry; Ministry of Education, Culture, Sport and Youth; Institution of research and innovation

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

4th Place at the First Pancyprian Cyber Security Marathon

IDEA Innovation Centre

Designed a multi-authentication system that using AI tried to verify the identity of a person using multiple traits of their Behavior.

'Best and Most Innovative Solution in Cyprus' in the WSA 2020 competition

WSA 2020

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

'Top 10 Green Businesses in Cyprus' in Climate Launchpad 2020 competition

Climate Launchpad 2020

Awarded for the development of a mobile application that aims to reduce the waste footprint of food providers by providing them with a way to sell edible products near expiration fast and effectively via an efficient matching with a potential interested party.

Third Place at the Pancyprian ACM ICPC (in Cyprus)

ACM ICPC

NeCo: Improving DINOv2's spatial representations in 19 GPU hours with Patch Neighbor Consistency

ICLR 2025

The paper proposes a new self-supervised learning method, NeCo (Patch Neighbor Consistency), which enhances pretrained representations by enforcing patch-level nearest neighbor consistency between a student and teacher model. This is done using a differentiable sorting method on top of pretrained models like DINOv2. The method achieves significant performance improvements with minimal computational cost (19 hours on a single GPU). It outperforms previous methods, setting new state-of-the-art results for semantic segmentation on several datasets, including ADE20k, Pascal VOC, and COCO-Things and COCO-Stuff.

[Re] Reproducibility Study of 'Label-Free Explainability for Unsupervised Models'

ReScience C

This work evaluates the reproducibility of the paper 'Label-Free Explainability for Unsupervised Models' by Crabbe and van der Schaar. The goal is to reproduce the paper's four main claims in a label-free setting: (1) feature importance scores reveal key features of a model's input, (2) example importance scores identify key training examples to explain a test example, (3) saliency map interpretability is difficult for disentangled VAEs, and (4) different pretext tasks have non-interchangeable representations.

Open Hummingbird Evaluation

Developed and published the Dense Nearest Neighbor Retrieval Evaluation (Balaˇzevi'c et al. "Towards In-contextScene Understanding") for testing the In-Context Learning Capabilities of vision encoders.

• Deep Learning
• Vision Encoders
• Self-Supervised Learning
• Nearest Neighbor Retrieval
• In-Context Learning
• Evaluation Metrics

Evaluating the Robustness of 3D Occupancy Prediction Models Under Noisy Data Conditions

This project focused on evaluating the robustness and reliability of State-of-the-Art 3D Occupancy Prediction Models under noisy or corrupted input data, simulated using synthetic data. The study highlighted the inadequacy of training models with data that does not reflect real-world scenarios. It demonstrated significant improvements in performance when applying appropriate data augmentations, achieving a 1-3% increase on noisy datasets and a 0.5-1% increase on clean datasets, further advancing state-of-the-art results.

• Computer Vision
• 3D Occupancy Prediction
• 3D Object Detection
• Synthetic Data
• Data Augmentation
• Robustness Evaluation
• Deep Learning

Iterative Image Refinement Using Socratic Models: Exploring Bias and Hallucination Effects

This project involved designing and implementing an iterative approach inspired by Socratic Models, which uses a pipeline of models to iteratively refine a prompt and generate improved images. The process involves leveraging a history of generated images and their refined captions (via a language model). The study revealed the bias and hallucination effects that emerge when connecting multiple models, highlighting challenges in model interactions and data generation.

• Deep Learning
• Image Generation
• Bias Detection
• Hallucination Detection
• Socratic Models
• Iterative Refinement
• Language Models