I'm trying to optimize a notoriously complex loss function, navigating a high-dimensional, non-convex landscape—one small step at a time. Probably, so are you!
- Choose your objective with caution—you drift toward whatever you optimize.
- Curate what you train on—you learn from what you keep feeding yourself.
- Set the pace—bold enough to move, small enough not to overshoot.
- Use momentum—keep going; let earlier progress carry you through flat stretches.
MEng in Electrical and Computer Engineering (Aristotle University of Thessaloniki, top 3% of class) and an MSc in Artificial Intelligence (magna cum laude, KU Leuven), plus two published papers on AutoML for anomaly detection and smart-meter security.
Since 2015: research associate at CERTH (EU cloud computing), software engineer at Veltio (supply-chain data pipelines for retailers like Sainsbury's), ML research engineer at ISSEL/AUTH (technical lead on an EU energy-monitoring project — NLP, event streaming), and since 2022, machine learning scientist at Expedia Group in London, working on large-scale ranking and retrieval with deep learning, LLMs, and multimodal methods.
Active, self-directed repos exploring ML/AI from first principles — most other side projects, coursework, and research code are archived.
- math-conceptual-map (site) — a 25-chapter conceptual map of mathematics for people who already know how to use it, tracing why each structure was invented, what it generalizes, and how logic, algebra, and probability converge on the same theorems on the way to a machine-learning bridge.
- modern-ai-systems-and-methods (site) — a 19-chapter field guide to modern AI, from supervised learning through agents and responsible AI, with real computed visualizations and explicit links between chapters.
- algorithm-visualizers (live demo) — 21 classic ML/CS algorithms, each a from-scratch NumPy implementation with an interactive Streamlit + Plotly walkthrough: clustering, dimensionality reduction, classification, deep learning building blocks, generative models, graph algorithms, probabilistic methods, and reinforcement learning.
- transformer-atlas (map) — a structured map of the transformer architecture, from the 2017 original through the attention variants, positional encodings, MoE routing, and inference tricks behind today's frontier LLMs, each with a minimal runnable PyTorch implementation.
- optimization-lab (docs) — from-scratch implementations of ~30 classical optimization algorithms behind a shared
Problem -> OptimizeResultinterface, a solver-arena comparison harness, and cross-domain applications in physics, economics, and ML. - loss-functions-lab (site) — where MSE, cross-entropy, hinge loss, RankNet, and DPO actually come from, derived from first principles (likelihood, MAP, decision theory, surrogate relaxations) rather than presented as a catalogue, closing with a reusable recipe for designing a new objective.
- computational-life-lab — an experimental artificial-life platform running the BFF primordial-soup experiment: self-modifying programs collide, recombine, and are studied for emergent self-replication and evolutionary dynamics, with no fitness function or hard-coded notion of "replicator."
Full CV here.