Senior Agentic AI Systems Engineer at Huawei
AI solution architect by trade: privacy-first RAG, LLM integrations, multi-agent orchestration, plus the open source tools underneath.
I am a Senior Agentic AI Systems Engineer at Huawei in Istanbul, and one of the founding members of a brand new agile team. What we build is automated code conversion, and the way we build it is AI-first development: agentic pipelines that carry a conversion job end to end, with multi-agent loop orchestration handling the parts a single pass cannot finish. My side of it is the LLM layer, the tools an agent acts through, and the software architecture that keeps a long running loop predictable once it is in production. Joining on day one means the team still gets to decide how those pieces fit together.
The four years before Huawei went into getting AI systems into production elsewhere. At .QIT I rebuilt Turkish customs clearance retrieval around RAG, fuzzy search and QLoRA fine-tuning, taking matching accuracy from 13% to a production-grade 96% entirely on-premise. At Rail-Flow GmbH I led AI and R&D: six engineers, three production systems, 5k+ documents a day on Docker and Kubernetes with LangChain and LangGraph. Earlier, at Simply Deliver, I built the PDF interpretation system that took roughly €100K a year out of manual processing.
The tools below are the same problem approached from the other side: running agents every day and needing to see what they actually did. They are local-first, dependency-light and open source. No server, no telemetry, nothing leaves your machine unless you decide it does.
Seven roles across four companies, from Oracle APEX tickets to a founding agentic systems team. Open it if you want the detail.
One portfolio site and three open source tools, all live and all deployed on Vercel.
The long version of this page: experience, enterprise projects, skills and education, in a terminal-styled interface that switches language on one click.
A roadmap, Kanban board and agent memory that run entirely on your disk. Four static pages rendered from JSON by a small CLI. No server, no build step, no network call.
Turn a coding session log into one self-contained HTML report: token spend per actor, three separate time measures, a parallel-agent timeline and every sub-agent task as a searchable row. The analysis runs in your browser, so the file never leaves it.
Istanbul, Turkey. Mail reaches me fastest. Everything I build in the open lives on GitHub.
