AI Engineer · LLM, RAG, fine-tuning · Python backend, React
I build practical AI systems in Python, from the model to a tested web application: question answering over documents with retrieval-augmented generation (RAG), multi-agent assistants, and small language models fine-tuned for a single task. I measure what I build: on a public benchmark in the RAG project, against simple baselines in the finance project. I also work on Turkish-language NLP.
Website · LinkedIn · Hugging Face · Kaggle · Email
| Project | What it does | Built with |
|---|---|---|
| Local RAG Q&A | Ask questions about your own documents, fully on your machine. Hybrid search (semantic + BM25) with a reranker, streamed answers with sources. Search measured on the SciFact benchmark: with the default settings, a correct passage is in the top 3 for 67% of 300 test claims. The test exposed a keyword-search bug, which I fixed. Measuring also led me to a better embedding model, which improves search by meaning on two separate datasets. | FastAPI, Ollama, ChromaDB |
| Multi-agent assistant | A supervisor agent sends each question to the agents it needs: researcher, analyst, editor. Web search, restricted Python execution, 93 unit tests. | LangGraph, Ollama, SearXNG |
| Zima | Health-advice assistant for people over 70. Synthetic dataset generation and LoRA fine-tuning of Qwen2.5-1.5B. | Unsloth, Transformers, Gradio |
| Finance sentiment | Labels a stock's news headlines with FinBERT, and measures it: on 2,388 labelled headlines FinBERT reaches macro F1 0.668, a simple model trained on the same kind of text 0.761. Tests, CI, one Docker image. | FastAPI, React, Docker |
| Datathon 2025 | Predicts how much each e-commerce session is worth from user events. | pandas, CatBoost |
| ALS recommender | Movie recommendations with collaborative filtering on Spark. | PySpark |
- TurkWeb-Edu Classifier scores the educational quality (0–5) of Turkish web text. It is a Turkish counterpart of the FineWeb-Edu classifier.
- TurkWeb-Edu Student is a 1.5B Qwen model that writes its reasoning in Turkish before it gives the score.
- Zima and its training dataset.
Python · PyTorch · Hugging Face Transformers · LangChain / LangGraph · Ollama · FastAPI · React · ChromaDB · scikit-learn · PySpark · Docker · GitHub Actions · AWS


