Onur Kansoy

I've built LLM systems from post-training research to deployed multi-agent products. My senior thesis, Smart Zebra, uses an end-to-end SFT and GRPO pipeline on Qwen2.5-3B-Instruct for logic-grid reasoning and achieved 66% accuracy on small puzzle instances.

At Research Ecosystems in Teknopark Izmir, I optimized proprietary AI models through bilingual query testing and engineered a multi-agent RAG system with LangGraph and PostgreSQL. I also developed GPT-Deep-Search, an autonomous OpenAI Agents SDK bot that discovers and aggregates academic profiles.

As an Undergraduate Researcher at Izmir Institute of Technology's Digital Humanities and AI Laboratory, I annotate and classify Ottoman Turkish documents through an AI-assisted workflow. I'm helping construct a ground-truth dataset for supervised fine-tuning of domain-specific language models intended for open sourcing on Hugging Face.

I've also shipped practical AI and product work, including a Terms of Service analysis Chrome extension, a multimodal bird identification pipeline, and a real-time mobile matchmaking app. Earlier, I designed game prototypes and helped win second place at the Kuleizmir Game Hackathon.