i'm a high school senior in ankara, türkiye, and a member of technical staff at freya (yc s25), where i work on speech: recognition, synthesis, and the training / data / eval pipelines that turn research models into production ones. i led the research and training of freyatts, the first turkish foundation text-to-speech model trained from scratch (183m parameters, tokenizer-free, non-autoregressive, flow matching). paper is number 5 below.
before that i was a research engineer at motionshark, building an ai tutor for the sat together with prep expert (the shark tank one), co-founded piyonist (an ai chess coach, in partnership with the turkish chess federation), and benchmarked a pile of llms for education at arf. before that i built an e-commerce integration platform in c# windows forms, which i mention only so you know i have suffered.
the question i think about most: what should i be doing right now to become the person who builds the next frontier lab? if you have an opinion on this, i genuinely want to hear it (email below).
the most recent paper is freyatts (see above). turkish is agglutinative, vowel-harmonic, and badly served by english-first speech models, so we trained one from nothing.
the second most recent one, ansatzu, is about automating variational quantum circuit design: search over the ansatz and the hyperparameters at the same time, while being honest about how many shots you can afford. qdpo, the one before it, tries to keep multi-reward reinforcement learning from collapsing by decoupling the rewards in a quantum-flavored way. superpositional gradient descent was my first conference paper and my first oral, at ieee qai 2025 in naples (see photo). echollama was the very first thing i put my name on: 3d-to-speech with multimodal ai.
naples, november 2025. ieee qai, first conference, first oral. the banner won the height contest.me, 2026. (yes, this is the good photo.)at üstünde ben, belli ki uzun zaman önce.
ansatzu: Automated Variational Quantum Circuit Design and Optimization with Shot-Aware Hyperparameter Search. Pamuk A.E., Yurtseven E. IEEE International Conference on Quantum Computing & Engineering (QCE) 2026, 4th International Workshop on Quantum Machine Learning: From Research to Practice.
QDPO: Quantum-Decoupled Policy Optimization for Stable Multi-Reward Reinforcement Learning. Pamuk A.E. IEEE World Congress on Computational Intelligence (WCCI) 2026, QNRL workshop.
i also write things up sometimes. the latest is a long post about gdpo experiments: training a turkish call-center agent against five rewards at once without reward collapse.
here are some things i believe in:
you don't need permission to do research. a laptop and an arxiv account are enough to start.
the fastest way to understand a thing is to train one from scratch.
turkish deserves frontier-quality models, and someone has to train them.
write things down, including the experiments that didn't work.
ask the dumb question early. it gets more expensive every week.
kaggle leaderboards are a great teacher and a terrible boss.
here are some things i like:
volleyball (i spike), chess (i lose to engines, which is partly why i built one), kaggle, math olympiads, the mit integration bee, the feynman lectures, mac mini clusters, karpathy tweets, "stack more layers", training runs that outlast my sleep schedule, tokenizer-free anything, people who reply to cold emails, ankara.
things i don't like:
tokenizers (see paper 5), reward collapse, windows forms.
