Semih Mutlu
I build small software products and local AI workflows.
I did product work at Madlen. I study CET at Boğaziçi, and I am a movie enjoyer.
Read more about me, see my projects, or inspect my AI workflow. I write at selambuarada.com.
How I work
I treat agents as infrastructure. Pi runs my day on a Fedora machine, in the same room as the work.
Pi. Main agent. DeepSeek as the cheap default model.
Herdr. Delegate board. Sub-tasks get panes; agents write back receipts.
Hermes. Previous always-on setup. Kept on disk as legacy.
Local verify-and-fix work runs autonomously. Anything public or irreversible — deploys, secrets, git history — stops for a human decision.
Movies Wrapped
Turns a Letterboxd export into a film-identity product: stats, share cards, a personal data story.
Madlen
At Madlen I did product QA and UX across an AI EdTech platform. One concrete change: tool-specific waiting states instead of a generic spinner. The pattern stayed. I also reviewed sessions, reported issues, and retested fixes. AI-assisted analysis, human judgment on what shipped.