I'm a Machine Learning Engineer who spent two years at Rozetka, Ukraine's largest e-commerce platform, building search, recommendation and auction systems used by millions of people.
Since then I've moved into interpretability research β most recently in the exploration phase of MATS 12.0, in Neel Nanda's mechanistic interpretability stream.
I'm open to research-focused roles in interpretability, LLM evaluation or ML engineering β based in Paris, remote, or happy to relocate, especially to the US.
I've always been a builder and a curious mind β from physics papers to tinkering with ML models, there's something genuinely satisfying about understanding how a system works and then pushing it further.
My background is deliberately varied: e-commerce recommendations, auction systems, anomaly detection, a startup, healthcare data. I like jumping between domains because the interesting problems rarely stay in one place β and the best ideas usually come from somewhere unexpected.
For a while I'd been spending more and more of my evenings on AI safety and LLM evaluation β now it's my full-time focus. The more capable models become, the more it matters to understand what's actually happening inside them, not just what they say on benchmarks. What I'm most curious about right now is the gap between what a model knows and what it tells you. It's a young field with a lot of open questions, which is exactly the kind of place I like to be.
And when I'm not poking at activations, I like making things go fast β most recently that meant 1st place at the GPU Mode IRL Hackathon in Paris, with an MoE inference engine doing on 8ΓH200.
And away from the screen, my newest hobby is building a drone from scratch β and learning to fly it, which turns out to be a whole different kind of debugging.
bridge-jspace β together with Zach Perlman, asking whether the hidden "bridge" step in two-hop reasoning (Amazon River β Brazil β Portuguese) flows through the part of the representation interpretability tools can read, or somewhere they can't see. The answer depends on where the model gets its answer: recalling a fact from memory leans on what the lens can read, while looking it up in the prompt takes a route the lens doesn't see at all.
secret-elicitation-audit β a model is fine-tuned to secretly know something about the user and deny it, and a benchmark scores how well interpretability tools can pull that secret out. A new probe seemed to beat the published methods, but the whole gain came from the layer it read, not from the probe β and the model grading every method turned out to have a favourite answer of its own. The lesson I keep coming back to: before crediting a method, check what else changed.
Smaller experiments β a refusal-direction replication, chain-of-thought control, and Deceptron, a scheming-detection pipeline built on ground truth you can execute β are in my repositories.
This is my playground for experimenting, learning, and building! You'll find interpretability experiments, ML projects, robotics simulations, physics-inspired projects, and probably some random stuff I'm just curious about. Each project is a learning adventure, and I love sharing that journey.
π¬ Always happy to connect with fellow builders, learners, and anyone curious about science and technology!





