Agents in the Sandbox: End-to-End Crash Bug Reproduction for Minecraft
E. Yapağcı, Y. A. S. Öztürk, E. Tüzün · Accepted to IEEE/ACM ASE 2025 · arXiv preprint ↗
AI engineer & researcher
I build with language models.
My work spans AI agents, speech recognition, document intelligence, and making models run faster. Here are a few things I’ve researched and built.
Bilkent CS ’25 · Research & engineering at Wharton, BILSEN, and DataBoss.
01 / Publications
E. Yapağcı, Y. A. S. Öztürk, E. Tüzün · Accepted to IEEE/ACM ASE 2025 · arXiv preprint ↗
A. S. Namlı, O. F. Celik, E. Yapağcı, S. S. Kozat · IEEE Transactions on Audio, Speech and Language Processing, Vol. 34 · DOI ↗
02 / Selected work
Open Turkish speech recognition that runs on a CPU. balbal-v2 transcribes 236× faster than real time on 8 CPU threads; balbal-v2-manas makes the fewest meaning-changing errors of any open model on Mu-Bench. Weights on Hugging Face ↗, code on GitHub ↗.
Experimental open speech recognition for Moroccan Darija, on a CPU. Lower error rates than Cohere Transcribe, Jais-ASR and MoulSot on Casablanca, MGB-5 and GigaSpeechBench MAR. Try it in the browser.
44 olympiad-level math problems to test how well AI models reason.
An optimized OCR pipeline with up to 3× faster CRAFT inference on large images.
A custom llama.cpp build for running DeepSeek V4 Flash on AMD Strix Halo, with optimized Vulkan kernels and speculative decoding.
03 / Get in touch
Want to talk about AI or something you’re building? Send me a note.
hello@erayy.com