Physics undergrad at Bilkent, self-taught software/AI engineer. Ankara, Turkey (UTC+3), remote.
I design and run AI agent systems and full-stack products end to end, by myself.
The stuff below isn't slides. It's deployed and running — mostly on Cloudflare, and most of it keeps running without me watching it. I build a lot, fast, and I ship.
My personal site, with more under the hood than a static page. Google OAuth login, an admin panel with live visitor analytics, and an LLM chat that routes to my own Claude subscription over a local bridge instead of paying per token, so each message costs me nothing. Built on Workers, D1, R2, Pages, and a Cloudflare Tunnel.
Two agent systems I built and keep running. papers-net drafts research papers through a master/director/worker/critic pipeline and checks its own output across four model families to catch hallucinations. karpathy is a personal autonomous agent: it sets its own goals, spins off work packets, and verifies what it finds. Both run on 5-minute systemd timers with SQLite state, on my own machine, no one watching them.
A free lecture-notes platform for my university, built with a few friends. 2,500+ materials indexed so far, plus a pipeline that auto-generates clean study notes from raw uploads. No ads, no paywall, and people actually use it. Live at thelecturenotes.com.
A Freqtrade bot, a Flask/Plotly dashboard, and conformal-bandit, a position sizer that uses conformal prediction to size bets by how confident the model actually is. I wrote it up as a methodology paper, including the honest result that the naive backtest oversold the edge, which is the part most writeups leave out.
A move-out / second-hand marketplace for Turkish university students. Student-email verification, a separate subdomain per campus, automated NSFW image pre-screening, bot protection, and the KVKK data-protection plumbing a real user platform actually needs — the unglamorous parts most side projects skip.
A spoken-Q&A tutor I built to practice technical interviews. It asks questions out loud with a local TTS voice, transcribes my answer with local Whisper, and a Claude-CLI grader scores the spoken answer on five axes, then reschedules weak concepts with spaced repetition (FSRS). FastAPI backend, runs fully on my own machine at $0 per session.
A 3Blue1Brown-style walkthrough of one of my cosmology papers — six animated scenes built with React-Three-Fiber and KaTeX.
A 3D avatar room that runs in the browser — Three.js front end fed by a Blender asset pipeline, embedded into my site.
An interactive storytelling web app — Vite + React, with the story logic and onboarding I cared about getting right.
Drafts I'm writing by hand and with my own agent tooling: a Pinney–Ermakov paper (sent to a professor), a smooth-classification paper, a Bohlin-cascade closed form, and a boundary-substrate cosmology paper. This is where the agent systems above and the physics feed each other.
A conflict-mapping and prediction platform: hash-committed forecasts you can't quietly edit after the fact, OSINT signal ingestion, and an AI editing copilot. On hold while I focus elsewhere.
I'm a physics student who got pulled into building software and didn't stop. I taught myself the stack and now I ship complete things on my own — usually faster than I should, occasionally before I've slept.
What I keep coming back to is systems that run themselves: agents that set their own goals, pipelines that check their own work, infrastructure that costs almost nothing to keep alive. I care about the boring parts too — auth, moderation, data protection, cost — because that's the difference between a demo and something real people use.