Hi, I’m Yelaman. Or just Yela.
I’m a research engineer based in Sydney. I spend most of my time thinking about generative and multimodal AI, building things, and trying to understand what actually works.
My background is a mix of AI research and software engineering, so I’m usually somewhere between experiments, model code, evaluation, and practical product work.
About
Right now I work as a Senior AI Engineer at StarRez. A lot of my current work is around integrating LLM features, designing evaluations, and using multimodal models for real product problems.
Outside of work, I like exploring model ideas and hard problems: reasoning tasks, custom language model architectures, post-training methods, ARC-style experiments, Kaggle competitions, and other things that feel interesting enough to try.
A few things I’ve worked on
Qurt
Qurt is a desktop AI coworker. I built it as a practical tool, but also as a way to explore model behaviour, interfaces, and workflows.
Research
Part of my PhD work has been around language data and synthetic data. I’ve also spent time working with modern open-source training stacks and post-training methods, including GRPO-style approaches, Unsloth, vLLM, Transformers, and related tooling.
But I also have spent time experimenting with ideas like byte-level energy-based language models, diffusion language models, JEPA-inspired language modelling, and other non-standard directions that I wanted to understand more deeply.
You can find my papers and citations on Google Scholar.
Seminars
I also organize seminars and reading club sessions in the DSML.KZ community. A couple of talks are here: Byte Latent Transformer and Energy-Based Transformers are Scalable Learners and Thinkers.
Reasoning and Kaggle
I like using competitions and open problems as a playground for ideas. That includes work around the AI Math Olympiad, ARC-style problems, and some side experiments like training BitNet-style models for chess.