About

Before language models, I worked on rockets. After studying rocket and spacecraft engineering at the Moscow Aviation Institute, I spent two years building satellites — useful, demanding work, but the part that kept pulling at me was the data and the models, not the hardware.

In 2021 I retrained into NLP and joined Sber as a data scientist. I started with classical ML, took on my own project a year in, and moved into a tech-lead role roughly eighteen months later. When LLMs became practical, I started pushing them beyond demos: into internal workflows, evaluation loops, and systems people actually had to rely on. One of those systems reached production early and became commercially useful, not just technically interesting.

My work today centers on internal agents — building new ones, improving existing ones, embedding them in the processes they need to live in, and running the R&D that supports both. The questions that pull me sit upstream of any specific product: how these agents behave when stitched together, how people end up working with them, and where the system starts to fail in interesting ways.

Outside work: music, badminton, and games — which I take seriously as art.

Currently

• Tuning LLM inference for an air-gapped environment — throughput, memory pressure, and observability that can't phone home.
• Collecting patterns for enterprise agents — evaluation loops, handoffs, fallbacks, and the boring parts that keep demos alive in production.
• Running small experiments on agent memory: what should persist, what should decay, and what should never be remembered.

Contact

osipchukevgeny@gmail.com
github.com/Osipchuk
x.com/EvgenyOsipchuk
linkedin.com/in/evgeny-osipchuk