About

I’m a year into an MSc in Computer Engineering at Karabük University, with a thesis on large-scale academic author disambiguation. Until Jul 2026 I was an AI Engineer at CodLab AI in Karabük.

At CodLab I was the sole engineer on an entity-resolution pipeline over bibliometric data — author disambiguation, designed and run in production, architected toward 85M+ articles (roughly 500M author mentions) on a single machine. The system is a two-stage Python ETL: PostgreSQL 15 with pgvector for 384-dim embeddings, Elasticsearch with a Turkish analyzer for fuzzy name search, SentenceTransformers, and RapidFuzz.

Matching runs through a multi-signal hierarchy — email, ORCID, ResearcherID, affiliation, the co-authorship graph, semantic embedding — with anti-snowball safeguards against false mega-clusters, and holds ~92% on a 200-case ground-truth benchmark with cluster-quality guards. A MinHash LSH block-then-cluster layer removed the Elasticsearch bottleneck entirely; an out-of-core memmap design keeps 85M-scale processing inside a fixed memory budget. Not everything worked: a ThreadPool Elasticsearch prefetch made things 35% slower, and a bucket filter cost 30 cases on the accuracy benchmark. The playground on this site reimplements the general technique with synthetic records.

Before that: at Mitsubishi Electric Innovation Center in Silicon Valley (Sep – Nov 2024) I optimized YOLO-based detection models for inference speed and accuracy; at DIATICS in Istanbul (May – Sep 2024) I did full-cycle computer vision — detection, classification, pose estimation, from data preparation and labeling through training and evaluation — served via FastAPI backed by PostgreSQL; at ArVis Technology (Jul – Sep 2023) I built a Unity game whose GPT-driven NPCs held unscripted conversations, with Azure text-to-speech and speech-to-text in a full voice loop; and at Deutsche Telekom (Jul – Aug 2023) I wrote Python automation for HR and back-office email workflows.