I work on the part of ML that comes after training: serving models fast, turning them into agents, and checking that they actually work. At Trendyol I ran model serving for 100M+ users with Triton and vLLM. On my own, I build AI agents with Google ADK and take them all the way to production.
Google Developer Expert, Cloud AI · GDG Istanbul co-organizer · Istanbul
Plus NVIDIA Triton · vLLM · Google ADK
- sdlc-agent-pipeline: agents that turn a GitHub issue into a tested pull request
- augury: a terminal reader that ranks new AI papers and posts for you
- tensorhane: a Turkish ML course with interactive demos
- claudemods: a community directory of Claude Code mods, plugins and skills (claudemods.chat)
- A single agent matched a planner + coder + reviewer team on 15 coding tasks (36 vs 34 of 45 runs solved) and cost less per run. sdlc-agent-pipeline
- At the same average rate, bursty traffic made vLLM's time to first token about 5× worse than steady traffic. vllm-metal-bench
- Splitting H100s between models with MIG gave about 1.75× more serving capacity. One H100, Many Models
Next: Retrieval Is All You Need (Almost) · DevFest İzmir 2026 · 5 December 2026
- Build Your Coding Agent with Google ADK · DevFest Kocaeli 2026 · workshop
- Building Trabzonspor's AI Analysis Room · DevFest Trabzon 2026 · workshop
- Deploying your agents on Google Cloud Agent Platform · Google Cloud Day Türkiye – Builder Connect · workshop
- One Framework to Deploy Them All! Meet NVIDIA Triton Inference Server · Oct 2026
- One H100, Many Models: What I Learned Sharing GPUs on Kubernetes · Sep 2026
- I Stopped Chasing the GDE Title. That’s When I Got It. · Sep 2026
- google/adk-web · Add light and dark theme toggle support to ADK Web UI
- google/adk-python · Fix parameter filtering for CrewAI tools that take **kwargs