ML Engineer focused on LLMs, RAG and deep learning. MSc in Computer Engineering at Bilecik Şeyh Edebali University, Türkiye.
I build ML systems end to end and measure every change against a baseline. I keep what wins and report what doesn't. Before AI, I spent 7 years in international trade, so I think about what a model is worth to the business, not only its score.
| Project | What it is | Result |
|---|---|---|
| enterprise-rag | Local RAG over SEC financial reports, evaluated on FinanceBench (150 analyst questions, 84 filings) with an LLM judge | 75.3% correct page in top 10 (BM25: 32%). A financial-statement boost lifted top-1 retrieval from 27% to 40% |
| minigpt-dostoevsky | GPT built from scratch in PyTorch, sized with Chinchilla scaling laws | 62.3% next-character accuracy vs 27.5% for a bigram baseline. Training past the Chinchilla budget cut test bpc by 33% |
| codebase-qa | Ask questions about a Python repo with local RAG | AST chunking 87.5% hit@5 vs 75% for fixed-size chunks |
| pneumonia-detection | Chest X-ray classification: CNN vs ResNet50 vs EfficientNet vs ViT | Fine-tuned ResNet50: 97.7% recall, 0.947 AUC |
| breast-cancer-streamlit | ML classifier deployed as an interactive web app | Live demo |
- Languages: Python, SQL, C++
- ML / DL: PyTorch, TensorFlow / Keras, scikit-learn, NumPy, Pandas
- LLMs & RAG: FAISS, BM25, embedding models, rerankers, Ollama, LLM-as-a-judge evaluation
- Apps: Streamlit
