Contact information

You can get this candidate's contact information from https://www.work.ua/resumes/18521982/

Work experience

No work experience.

Education

Secondary

Additional education and certificates

Mate Academy

2025, 12 місяців

AI Engineer Core Track: LLM Engineering, RAG, QLoRA, Agents — Udemy (Ed Donner)

2026
Certificate

Knowledge and skills

• Python
• Git
• HTML
• SQL
• Знання принципів ООП
• GitHub
• CSS
• PostgreSQL
• Docker
• Django
• MySQL
• JavaScript
• FastAPI
• SQLite
• Django Rest Framework
• REST
• Flask
• LLM
• RAG
• PyTorch
• Hugging Face
• Fine-tuning
• Ollama
• LangChain
• Pydantic
• Pytest
• Chroma

Language proficiencies

English — above average

Additional information

Junior AI / LLM engineer looking for a first role: remote, or office/hybrid in Barcelona. English B2, Ukrainian native.

I finished Ed Donner's LLM Engineering course by building one project through it: Spanish Tutor AI (github.com/MrEug3n1o/spanish-tutor-ai), running entirely on local models via Ollama.

- Tool calling (verb conjugation, flashcard decks), streaming Gradio UI, Whisper voice input.
- RAG: BGE + Chroma, query rewriting, cross-encoder reranker; MRR 0.648 → 0.891 on a 36-question eval set.
- QLoRA of Llama 3.2 3B on MLX: el/la for 98.6% of 275 unseen nouns (llama3.2: 70%), on par with qwen2.5:14b in a fraction of the time.
- Product pricer (course labs, https://github.com/MrEug3n1o/llm_engineering): error $190 → $56.25, ahead of GPT-4.1 Nano ($62.51).
- Code assistant with an AST check that docstrings never change code; pytest suites repaired in a loop.
- Coach agent: a tool-calling agent on qwen2.5:14b with spaced-repetition memory that picks the topic, writes a deck past the article checker, writes a quiz and sends the lesson; 12/12 eval scenarios completed with the right topic (llama3.2: 0/12).

Two more measured projects:
- Smart Call Tracker (github.com/MrEug3n1o/smart-call-tracker): call-centre QA with Whisper and an LLM, fully local or with Gemini. 96.2% of QA labels right on held-out test calls, WER 13.2% on phone audio, 68 tests in CI. Labels are mine; agreement with a human reviewer isn't measured yet.
- SphereID (github.com/MrEug3n1o/sports_classifier_ml): EfficientNet-B0 fine-tuned in PyTorch, served as ONNX. 89.5% on unique test photos vs 84.3% for my old model, 13× faster on CPU.

Before that: Python backends with Django REST + PostgreSQL, FastAPI, DuckDB + Streamlit analytics in Docker.

UPML graduate, regional informatics olympiads 2024–2025, Mate academy Python Developer certificate, SULICOM agentic-engineering workshops. No commercial experience yet.
GitHub: github.com/MrEug3n1o