About
I'm a researcher working on Federated Learning and Computer Vision. I'm wrapping up my PhD at City St George's, University of London, where I design architectures and aggregation methods for decentralized training under data heterogeneity.
Before London, I studied Electrical and Computer Engineering in Thessaloniki, then moved to Imperial College for my MSc in AI. Somewhere in between, I worked at Equideum Health on contribution evaluation for federated data markets.
I work on federated learning for distributed, heterogeneous data. My research focuses on making decentralized training practical when clients hold data that varies in distribution, modality, or imaging characteristics — the norm in real-world medical and multimodal settings.
Architectures and Training Pipelines for FL
I study how architecture choice, weight initialization, and aggregation method interact in federated visual classification. My benchmarking work shows these elements must be chosen jointly — ImageNet pretraining helps, but self-supervised pretraining on domain-relevant data can match or exceed it; normalization-free networks and transformers each have distinct failure modes under heterogeneity.
Aggregation Under Heterogeneity
I design aggregation methods that adapt to client-level training dynamics rather than treating all participants uniformly. FedCLAM derives per-client momentum and dampening from local validation progress, combined with a foreground intensity matching loss that handles scanner-specific brightness and contrast biases.
Federated Multimodal Learning
I'm building benchmarks and methods for federated fine-tuning of multimodal large language models, where clients may hold different modalities entirely — missing images, text-only data, or mixtures. This introduces a new axis of heterogeneity beyond the label skew typically studied.
Federated Reinforcement Learning
My MSc thesis developed a framework formalizing how multiple data owners can collaboratively train RL agents without sharing raw trajectories, covering privacy-preserving representations, aggregation strategies, and evaluation under client heterogeneity. I applied this to sepsis treatment across real ICU partitions in MIMIC-III and released FeRaL, an open-source library for FRL experimentation. As autonomous agents become increasingly deployed and begin interacting with one another, the intersection of federated learning and multi-agent RL feels ripe for revisiting.
Earlier Work
My diploma thesis at Aristotle University focused on learning under label noise in image classification — an early encounter with the challenge of training on imperfect supervision that has informed my subsequent work on robustness under data heterogeneity.
Education
PhD in Computer Science
City St George's, University of London
MSc in Computing (AI & ML) — Distinction
Imperial College London · Bodossaki Scholar
MEng in Electrical & Computer Engineering — Top 5%
Aristotle University of Thessaloniki
When I'm not working, I'm usually maintaining a 20-year-old sports car, hiking, or at the range.