littlestronomer
Per aspera ad astra
I'm an AI engineering student at Istanbul Technical University, working on ML systems, speech AI and fast inference.
Learning a disc inside a ring
This is a real neural network, training in your browser as you read. It only knows where each dot is, and learns to color the map: blue where blue dots belong, orange where orange ones do. Once it gets them right, it moves on to a harder puzzle.
Each square is one neuron, showing which parts of the map make it fire. The lines are the weights between them: blue ones add, orange ones subtract, and brighter ones count for more.
I'm building Waifu SpellForge, a card game about anime-style spellcasters. An LLM writes each spell card and a text-to-image model paints it.
It runs on vLLM, Ollama and Stable Diffusion with quantized settings, so it fits on GPUs with little memory and deploys to Hugging Face Spaces. The game rules are still on the drawing board.
We rebuilt the Agenthon ABIDES market in C++ to match the original event for event, and only then made it fast.
I built the preprocessing pipeline for 3.7 TB+ of speech audio, running 2 to 4 times faster than real time, and fixed broken metadata in 23% of a 500K-sample dataset. I fine-tuned XTTS for a single voice, improving both similarity and speed, and set up CosyVoice2 fine-tuning for multi-speaker voices.
I compared BERTurk and mT5 for predicting customer satisfaction from Turkish banking text, and tested Vision Transformers on financial documents.
An agent with 45 tools that answers questions about the Turkish economy from 13 million rows of central bank and banking data, built with two teammates for the KKB Hackathon. It took 1st place.
Separate CUDA paths for decoding and prefill, 1.07× and 5.33× faster than FlashInfer. It placed 4th in the NVIDIA MLSys Competition.
My entry to the AMP Challenge 2027, an international competition on generative AI for new antimicrobial peptides. Two small autoregressive models, the main one with 10.7 million parameters, generate a library of 50,000 peptides, and a selector ranks the top 100.
Real-time transcription that tells speakers apart, sped up with quantization and PyTorch Dynamo.
Finds idioms in Turkish and Italian sentences with BERT and RoBERTa, at an F1 of 0.92.
Removes chosen people, objects and colors from a diffusion model, following Forget-Me-Not.
My merged pull request lets this 2.6K-star open-source project run fully offline.
A data agent for Turkish finance, with Ahmet Selim Özel and Alihan Esen. See the project
We turned the Turkish school curriculum into short narrated videos.
Training only Llama 3's last 8 layers beat both LoRA and full fine-tuning.
I'm currently ranked 3,647th on Kaggle. See my Kaggle profile
I speak and English, and I'm learning Japanese and Chinese.
I'm open to research collaborations, internships and ML systems roles. Email is the fastest way to reach me.