Skip to content
View georgebon5's full-sized avatar

Highlights

  • Pro

Block or report georgebon5

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
georgebon5/README.md

George Bonovas

Software engineer working on LLM inference and backend systems, based in Athens. Final-year CS at NKUA, writing my thesis on KV cache quantization.
I work mostly on backend systems and AI tooling, and I like owning the whole thing: data layer, API, and the ops around it.

Website LinkedIn Email

Open to internships and junior roles in AI engineering or backend. Remote or hybrid.

Projects

LLM inference engine written from scratch in C++20. No PyTorch, no ggml, no BLAS.

  • Full transformer stack (GQA, RoPE, SwiGLU, KV cache) on hand-written NEON SIMD kernels
  • Runs SmolLM2 135M to 1.7B on Apple Silicon / ARM CPUs at 10–61 tok/s
  • Loads BF16 weights straight into INT8, so the 1.7B model fits in under 2 GB of RAM
  • Web UI with live token streaming, a per-token top-5 probability panel, and FP32 / INT8 / INT4 side by side

C++20 NEON SafeTensors POSIX sockets CMake

Backend where organisations upload documents and chat with them over RAG. Each tenant's data is fully isolated, and new tenants are onboarded without touching existing ones.

  • ~13k LOC, 55 endpoints, 330 tests
  • Celery + Redis for ingestion jobs, WebSocket streaming for LLM responses
  • JWT auth, RBAC, Kubernetes manifests and CI/CD (clone to running in one command)

FastAPI ChromaDB PostgreSQL Celery Redis Docker Kubernetes

Three-way clarity classification of politicians' answers on the QEvasion dataset. I ran four approaches on the same split: TF-IDF / Word2Vec baselines, fine-tuned BERT / DistilBERT / DeBERTa-v3 (plain PyTorch, no Trainer), Qwen2.5-3B/7B prompting (zero-shot, few-shot, CoT), and a four-agent pipeline on a 0.5B model.

The final submission, a 6-encoder ensemble with one round of pseudo-labelling, placed 2nd on the private Kaggle leaderboard (macro F1 0.67).

PyTorch HuggingFace Qwen2.5 scikit-learn

Churn prediction on 10k bank customers, built as a full pipeline rather than a notebook. DVC for data, MLflow for experiments (gradient boosting won, ROC-AUC 0.730), FastAPI serving, Evidently AI for drift detection, GitHub Actions for CI/CD.

MLflow DVC FastAPI Docker Evidently AI

Next Jr by NBG · TechBiz Hackathon

🏆 1st in the National Bank of Greece track, 3rd overall. A youth banking app with AI features, built and pitched to NBG in under 24 hours: two codebases, GDPR compliance, Claude API integration. I did the frontend, the API design, and the pitch.

Next.js TypeScript FastAPI Claude API

Coursework
Project
Database Systems C, B+ trees, file organisation, indexing
Operating Systems C, processes, memory, scheduling
RSA C, encryption from scratch
OpenGL shadow renderer C++, shadow mapping, Phong lighting
Pacman Python, search and adversarial agents

All in programming-projects.

Stack

Languages   Python, C++, C, SQL, TypeScript
ML   PyTorch, HuggingFace, scikit-learn, ChromaDB, Anthropic SDK, MLflow, DVC, Evidently AI
Backend   FastAPI, Flask, PostgreSQL, Redis, Celery, WebSockets
Frontend   Next.js, React
Infra   Docker, Kubernetes, GitHub Actions

Pinned Loading

  1. MLOps-Pipeline MLOps-Pipeline Public

    End-to-end MLOps pipeline for bank churn prediction: DVC, MLflow model registry, FastAPI serving, Docker, CI/CD and Evidently drift monitoring.

    Jupyter Notebook

  2. Multi-Tenant-AI-Document-Pipeline Multi-Tenant-AI-Document-Pipeline Public

    Production-ready Multi-Tenant AI RAG System — upload documents and chat with them using Claude/GPT. Built with FastAPI,ChromaDB, PostgreSQL, Celery.

    Python

  3. nbgtechbliz nbgtechbliz Public

    Next Jr. by NBG - youth banking MVP

    TypeScript

  4. minillm-cpp minillm-cpp Public

    From-scratch C++ LLM inference engine — runs SmolLM2 (135M–1.7B) on CPU with hand-written NEON SIMD, INT8 quantization, and a live web UI.

    C++

  5. political-response-clarity political-response-clarity Public

    3-way classification of political response clarity (SemEval 2026 Task 6). TF-IDF, fine-tuned BERT/DeBERTa, Qwen2.5 prompting, D3-agentic pipeline — 2nd on Kaggle leaderboard.

    Python