A little about me

Learning to represent,
reason, and act.

I'm Semen, a research engineer interested in how learned representations of the world can support reasoning, planning, and decision-making.

I currently lead tool-use post-training for GigaChat Omni at Sber, working on cold-start data, evaluations, and reinforcement learning. Previously, I started the team's VLM effort at Yandex Autonomous Vehicles, taking models from raw sensor data to on-vehicle testing.

I'm completing an M.Sc. in Deep Learning and Generative Artificial Intelligence at ITMO University, where I also studied Applied Mathematics and Computer Science.

How can we learn predictive representations that help models reason about what comes next?

Research

My interests sit at the intersection of world models, self-supervised representation learning, latent dynamics, and model-based reinforcement learning.

Predictive representationsResearch prototype · Yandex

Planning in latent space

I explored JEPA-based trajectory planning through sequential prediction of future latent states. A compact model with a ViT-B backbone and a roughly 20M-parameter latent predictor outperformed a substantially larger autoregressive planning baseline on selected driving scenarios.

Multimodal reasoningOngoing · ITMO & GigaChat

Allocating computation to reasoning

I'm studying variable-length latent and textual reasoning in vision–language models. I implemented up-front budget prediction and learned stop heads as alternative ways to control reasoning length.

I reproduced LVR, HyLaR, and LUT; current experiments use Qwen3.5 and Qwen3.8 (27B).

Data & perceptionBachelor's thesis · 2025

Finding data for rare road objects

I built a VLM-based pipeline for mining perception errors and selecting examples for detector fine-tuning, evaluating Yandex VLM, Qwen2.5-VL, and InternVL2.5 on 50,000 images. For the object-counting component, prompt tuning reduced MSE by 31% (15.55 → 10.73) relative to the same 35B baseline without adaptation.

Experience

2026 — now

Sber / GigaChat Omni

Lead, Reasoning and Tool Use

I initiated and own the tool-use training effort as its sole technical contributor: web search, thinking with images, calculator use, and MCP tools. The work includes evaluations with tool execution and GRPO/CISPO recipes with judge-based and algorithmic rewards.

2024 — 2026

Yandex / Autonomous Vehicles

Software Engineer, ML R&D

I built the full model release pipeline, from mining raw logs and preprocessing video/LiDAR to training, evaluation, simulator testing, and vehicle deployment. A deployed VLM planner predicted 50-step trajectories over five seconds with 200 ms inference latency and passed test-track safety and maneuver tests.

Training and inference workloads on hundreds of GPUs; data tooling for archives containing hundreds of petabytes of raw logs.

2021 — 2024

T-Bank / Core Technologies

Software Engineer

I developed internal ML services and software tools for bank employees.

Teaching

Algorithms and Data Structures

Course creator · T-Bank

Created the course for university students across three spring offerings.

Spring 2024
2025 & 2026

Programming at ITMO

Teaching assistant · Seminars

Introduction to Programming (Fall 2024) and Programming Paradigms (Spring 2025), courses led by Georgiy Korneev.

2024 — 2025

Publications & talks