Head of LLM Project · Pre-Training · Alignment · Applied LLM
I lead pre-training, alignment and applied LLM teams behind the GEN-T family of open-source Russian-language large language models — T-Lite, T-Pro, T-Pro 2.0 and T-Pro 2.1. Our models ship to millions of users.
I'm an ML engineer in Moscow. I run the LLM project — three teams, each with its own lead, covering the full post-training lifecycle: pre-training, alignment (SFT + RLHF) and applied LLM. Our models power internal copilots, customer support agents and an assistant ecosystem used by millions, and we release them as open weights to the Russian-speaking ML community.
Since 2018 I've worked on NLP and LLM engineering — first as an NLP engineer, then leading AI support from 2021, and heading the LLM project since 2023. Before that I was an NLP researcher at DeepPavlov, Mikhail Burtsev's lab at MIPT.
Shipping the next generation of the GEN-T family and scaling the post-training team. My day-to-day is research direction, training-recipe decisions, hiring and org design — I don't write production code anymore, but I stay deep in the tradeoffs.
I hold Israeli citizenship and am in the process of obtaining a UK Global Talent visa. I'm open to leadership roles at frontier AI labs in London, Tel Aviv or the Bay Area.
Head of LLM Project
I lead the group behind the GEN-T model family and the LLM stack powering our products.
Tech Lead / Team Lead — AI Support
I built the AI stack for customer support — intent routing, dialogue systems, agent assistance.
NLP Engineer → Senior NLP Engineer
I joined the NLP team and worked on classification, information extraction and dialogue components for production systems.
NLP Researcher — DeepPavlov (MIPT Neural Networks and Deep Learning Lab)
I contributed to DeepPavlov, an open-source library for dialogue systems and chatbots developed at Mikhail Burtsev's lab at MIPT.
T-Pro 2.1 (33B) and T-Lite 2.1 (8B)
T-Pro 2.0 (33B)
HF model · Technical report. See Publications for the EACL 2026 paper.
T-Pro (32B) and T-Lite (7B/8B)
Our first public GEN-T release. Set a new bar for open Russian-language LLMs on industry benchmarks. Technical report.
Open-source AI/ML ecosystem
How our team builds and releases open-weights LLMs. Overview.
T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground
We present T-pro 2.0, an open-weight Russian LLM for hybrid reasoning and efficient inference. Key contributions: a Cyrillic-dense tokenizer that cuts average tokens-per-word from 3.12 to 2.38 on Russian text, and an adapted EAGLE speculative-decoding pipeline delivering up to 2× inference throughput.
Building a Strong Base LLM from Open Models
How we adapt open base models (Qwen, Llama) into strong Russian-language LLMs like T-Lite and T-Pro — architecture choices, tokenizer, continued pre-training and alignment. Watch on YouTube.
Where to get data for LLM pre-training
My talk on how we source and curate pre-training data for Russian LLMs. Watch on YouTube.
Moscow Institute of Physics and Technology (MIPT)
Moscow Institute of Physics and Technology (MIPT)