Hello, I'm

Dmitrii Kuzmin

NLP/ML Engineer · NLP Researcher

Fine-tuning, inference, and tokenizer research for large language models.

Something went wrong

About Me

NLP/ML engineer with experience in both research and product work on large language models. Specialized in fine-tuning (SFT, LoRA/QLoRA) and inference of open-weight models (Qwen2.5-VL, Kimi2.6), building agentic pipelines (LangGraph, LangChain, LiteLLM), and research on tokenization and LLM uncertainty estimation.

Publications include AI Journey 2025 (accepted, poster) and two A* conference papers currently under review. Hands-on with GPU clusters (NVIDIA, MetaX), CUDA, Docker, and Git — comfortable moving from a notebook to a production serving stack.

Something went wrong

My Skills

Balanced toolkit across research, engineering, and deployment for large language models.

Download CV

LLM / NLP

  • PyTorch, Hugging Face Transformers
  • PEFT (LoRA / QLoRA), SFT
  • Tokenization (BPE, SentencePiece)
  • vLLM, Triton inference

Agentic & RAG

  • LangChain, LangGraph
  • LiteLLM routing across providers
  • Pydantic data models
  • Qdrant vector store

MLOps / Infra

  • CUDA, Docker, Git, Linux, Bash
  • Weights & Biases experiment tracking
  • Distributed inference on NVIDIA & MetaX GPUs

Data & Backend

  • Python, NumPy, Pandas, scikit-learn
  • SQL query optimization
  • FastAPI services

Languages

  • English (C1, confirmed)
  • Russian (native)

Soft Skills

  • Flexibility
  • Responsibility
  • Enthusiasm

Experience

Hands-on roles delivering large language model research, productionization, and tooling.

NLP Engineer · Neural Systems and Deep Learning Lab, MIPT

Moscow, Russia · May 2025 – present
  • Maintain and scale LLM inference across 3 clusters of 17 GPUs (NVIDIA A100/H100 and MetaX), serving 6 models from 14B Dense to 1T MoE at 150k+ requests/day; reduced p95 latency from 3000 ms to 2500 ms (17%).
  • Lead research on uncertainty estimation in the reasoning of large language models.
  • Designed and deployed a LangGraph agentic pipeline that generates Selenium tests from natural-language descriptions.
  • Designed the architecture of a co-pilot for automating call-center operator workflows.
  • Designed an LLM-uncertainty evaluation benchmark covering 40+ models across 4 domains.
  • Launched a recurring LLM-eval (LLM-as-a-Judge, benchmark criteria) pipeline on MetaX GPUs.

Visiting Researcher, NLP · MBZUAI

Abu Dhabi, UAE (remote) · Jun 2025 – present
  • Lead research on adapting LLM tokenizers to Russian on Qwen2.5-1.5B-Instruct (1000+ GPU-hours), achieving a 17% reduction in fertility relative to the baseline tokenizer.
  • Proposed and implemented an alternative tokenization method.
  • Coordinated experiments with collaborating research teams.

Middle NLP Engineer · Center for Applied AI, Skolkovo

Moscow, Russia · Feb 2025 – May 2025
  • Fine-tuned Qwen2.5-VL (LoRA, r=16, 3 epochs) on a 200-drawing corpus for object detection: mAP@0.5 = 0.6, mAP@[.5,.95] = 0.35.
  • Assembled an end-to-end pipeline for object detection in drawings (preprocess → inference → postprocess → report) with 87% class coverage.
  • Selected and A/B tested prompts for generating remarks on drawings.

Research Assistant, NLP · Higher School of Economics

Moscow, Russia · Jun 2024 – May 2025
  • Fine-tuned Llama-3-8B-Instruct (SFT, 300 GPU-hours); achieved an 11.6% improvement on the overall MERA benchmark score.
  • Developed a BPE tokenizer for Russian on an 80 GB corpus; reduced the average number of fragments per word from 3.86 to 1.5 (61% reduction).
  • Implemented methods for manipulating existing tokenizers (token deletion / addition / merging) — released as the open-source library TokenizerChanger.
  • Created a grammar benchmark for Russian, used in the AI Journey 2025 publication.

ML / Backend Engineer · Moscow Aviation Institute

Moscow, Russia · Jul 2023 – Oct 2023
  • Developed a sentence topic classification model with 10+ classes.
  • Optimized SQL queries and database schema: p95 latency reduced from ~1500 ms to ~800 ms (47% reduction).

NLP Engineer (Internship) · Innopolis University

Innopolis, Russia · Jun 2023 – Jul 2023
  • Fine-tuned RuBERT for sentiment analysis of YouTube comments on a ~10k example dataset; F1 = 0.82.

My Projects

Publications

Researching tokenizer adaptation and cost-efficient strategies for large language models.

A Multi-Aspect Evaluation of Tokenizer Adaptation Methods for LLM on Russian

AI Journey 2025 · Co-author · 2024 – 2025 · Accepted (poster)

Systematic comparison of 4 BPE tokenizer adaptation methods for Russian-language LLMs across 3 benchmarks (MERA, MMLU, TA bench) and additional statistical methods. The best method reduces fragmentation from 3.7 to 2.34 (−36.7%).

mathnet.ru

Mitigating the Impact of Glitch Tokens via Targeted Retokenization

AACL 2026 · Co-author · 2025 – 2026 · Under review

An inference-time retokenization method that removes under-trained ("glitch") tokens based on embedding norm/entropy; no model fine-tuning required. Evaluated on 6 models (1.5B – 32B) on MMLU and WMT: +7% improvement in robustness to glitch prompts.

TokenSubstitution: Cost-Efficient Method of Language Adaptation Based on Token "Trained-ness"

EMNLP 2026 · First author · 2025 – 2026 · Under review

A method for adapting an LLM to a new language by replacing weakly trained tokens with tokens of the target language, without full embedding fine-tuning. −15% inference latency while surpassing the original model on the MERA benchmark.

Education

Grounded in data analysis, software systems, and human-centered AI design.

B.S. in Data Analysis and Artificial Intelligence

Innopolis University · 2022 – 2026
  • Core courses: Software Systems Analysis & Design, Human-AI Interaction Design, Mathematical Analysis.
  • Focus on applied machine learning, large language models, and productized AI systems.

Extracurricular Activity

Tutor · Innopolis University · Sep 2023 – Jan 2024
  • Helped first-year students adapt to academic and cultural workflows.
  • Organized extracurricular events to foster community and peer learning.

Contact Me

Open to research collaborations and engineering roles in NLP, LLM fine-tuning, and tokenizer work.

Created by Dmitrii Kuzmin