Ozan Berk Gültegin

I've built an end-to-end honey-bee behavior recognition pipeline for my MSc thesis at Vilnius Gediminas Technical University, combining YOLOv11, SAHI, ByteTracker, DINOv2, and a temporal Transformer. The system achieved 0.853 mAP@0.5:0.95 and macro F1 of 0.96, a 31% relative improvement over the prior YOLOv8 baseline.

I've also architected a hybrid arXiv RAG pipeline that ingests papers through the arXiv API, PyMuPDF, LangChain, Qdrant, and PostgreSQL. Its dense retrieval, BM25, and RRF fusion approach delivered sub-100ms latency and achieved 0.95 answer relevancy on a curated benchmark.

As a solo builder, I designed and shipped Telegrad, a production Telegram advertising marketplace with a Next.js campaign workflow, Telegraf bot, billing pipeline, CPM auction engine, fraud detection, moderation tooling, and advertiser-publisher workflows.

Previously, I improved hospital-data ETL processing by 17% at SKSPro and modernized C# and SQL analytics systems at Tarnet. I'm an MSc Artificial Intelligence Systems valedictorian with a 9.7/10 GPA and a Software Engineering background.