About

I'm a Computer Science student in Athens, mostly working on backend systems and AI tooling. The projects I enjoy the most are the ones where I own the whole pipeline — from the data layer up to the API and the ops around it.

Over the last year I've built a production-grade multi-tenant RAG system (55 endpoints, 330 tests, Kubernetes), an end-to-end MLOps pipeline with drift detection, and won first place in a national banking hackathon.

For my thesis I'm now researching KV cache quantization for efficient LLM inference — a natural continuation of the inference engine work in minillm.cpp, this time at the activation side rather than the weights.

I'm currently looking for an internship where I can keep working close to the metal — distributed systems, data pipelines, or applied AI. Remote or hybrid, anywhere.

Selected Work

01

minillm.cpp

C++20NEON SIMDSafeTensorsPOSIX socketsCMake

From-scratch LLM inference engine — no PyTorch, no ggml, no BLAS. Implements the full transformer stack (GQA, RoPE, SwiGLU, KV cache) with hand-written NEON kernels. Runs SmolLM2 (135M–1.7B) at 10–61 tok/s on CPU. Direct BF16→INT8 loading keeps the 1.7B model under 2 GB RAM. Web UI with live token streaming, per-token top-5 probability panel, and FP32/INT8/INT4 side-by-side comparison mode.

02

Response Clarity Classification — SemEval 2026 Task 6

PythonPyTorchHuggingFaceQwen2.5scikit-learn

Three-way classification of political response clarity on the QEvasion dataset. Ran four approaches end-to-end on the same split — TF-IDF/Word2Vec baselines, fine-tuned BERT/DistilBERT/DeBERTa-v3, Qwen2.5-3B/7B prompting (zero/few-shot/CoT), and a four-agent D3-agentic pipeline on a 0.5B backbone. Final submission is a 6-encoder ensemble with pseudo-labelling — 2nd place on the private Kaggle leaderboard (macro F1 0.67).

03

Multi-Tenant AI RAG System

FastAPIChromaDBCeleryRedisKubernetes

Multi-tenant backend where users upload docs and chat with them. ~13k lines — 55 endpoints, 330 tests. Celery and Redis for background jobs, WebSocket streaming for LLM responses, full RBAC, Kubernetes deploys.

04

End-to-End MLOps Pipeline

PythonFastAPIDockerMLflowDVC

ML pipeline predicting bank customer churn on 10k samples. Three models tested, best ROC-AUC 0.730 with gradient boosting. Evidently AI for drift detection, GitHub Actions for everything else.

Coursework

PacmanPython — AI agents & game logic
Database SystemsC — B+ trees, file organization, indexing
Operating SystemsC — processes, memory, scheduling
RSAC — encryption from scratch
opengl-shadow-rendererC++ — 3D renderer, shadow mapping, Phong lighting
Full repo →

Skills

Languages

PythonC++SQLTypeScriptJavaScript

ML / AI

PyTorchScikit-learnChromaDBAnthropic SDKEvidently AIRAGLLM integration

Backend

FastAPIFlaskPostgreSQLRedisCeleryWebSockets

Frontend

Next.jsReactTypeScript

Infrastructure

DockerKubernetesGitHub ActionsMLflowDVC

Awards

TechBiz Hackathon — NBG Challenge

1st NBG / 3rd Overall

National Bank of Greece

  • Competed in the National Bank of Greece track — built a youth banking app with AI features.
  • Two codebases, GDPR compliance, REST API mockups — all in under 24 hours.

Education

National and Kapodistrian University of Athens

Expected 2027

BSc Computer Science — GPA 8.2 / 10.0 · Athens, Greece

Coursework: Algorithms & Complexity, Machine Learning, Operating Systems, Distributed Systems, Graph Theory.

Workshop: Innovation, Entrepreneurship & Deep Tech

Nov 2025

EIT Higher Education Initiative

Contact

Open to intern and junior roles in AI engineering or backend — hybrid or remote.

If you have a question, an opportunity, or just want to say hi — my inbox is always open.