Data Scientist

Work Experience

Skills

Background

Skills

Machine learning:

Large language models:

Prompt engineering
Fine-tuning

Deep learning (PyTorch, HuggingFace, Ollama):

Transformers
Use with limited computational resources: quantization, gradient accumulation, PEFT, LoRA.

Classical ML (sklearn, xgboost):

Regression: linear, nonlinear, regularization (lasso, ridge)
Classification: logistic regression, KNN, SVM
Clustering
Tree-based methods: random forest, boosting

Cross-validation, bootstrap.
Hyperparameter optimization (Optuna)
Feature transformation (PCA, SVD)
Ensembles, pipelines

Speech to text and text to speech
Natural language processing (NLTK, SpaCy, Stanza):

Machine translation
Machine translation quality assessment
Tokenization
Named entity recognition
Text classification

Statistics:

Experiment design
Hypothesis testing
Statistical power and sample size estimation
ANOVA
Multiple testing
Bootstrap

Exploratory data analysis
Python: Pandas, NumPy, Matplotlib, SciPy, StatsModels, Jupyter Lab, etc.

Tools: Jupyter Lab/Jupyter Notebook for experiments; PyCharm/Visual Studio Code for development.
FastAPI
Docker
Git
SQL
Databases: PostgreSQL, Clickhouse
BI: DataLens, Superset

Back to top

Work Experience

April 2025 — present. AI Engineer in Freedom Finance Technologies (fintech)

Developing a corporate AI assistant with a microservice architecture (FastAPI, Docker Compose).

My products included the following:

Automatic speech recognition and processing services.

Problems solved:

Dialog compilation from transcripts produced from a record with each speaker in their own channel. Correct processing of simultaneous speech and interruptions.
Prevention of hallucinations caused by long pauses in some languages (Greek)
Improved detection of gaps in transcripts

Prompt engineering:

Re-engineering of phone call analysis prompts with a substantial reduction of false positives

June 2023 — March 2025. Language Researcher, Expert Center, AWATERA, a large language service provider

The Expert Center is responsible for research, new technologies and other non-routine tasks. My products included the following:

Corporate speech recognition service:

Accessisble to all employees of the company via a web-interface
FastAPI backend
Options for locan and remote deployment, including in Docker containers
Transcript segmentation by sentences or by time
Gap and hallucination detection
Transcript editing by an LLM
Various output formats
Transcript delivery by email
New features on user requests
Mantenance: version control (Git), logging and bug fixing
Documentation, user and developer training

Text domain classifier

Two tiers of classification: upper(industrial, medical, etc.) and lower (for medicine: pharma, clinical trials, etc.). Dedicated classifier for each language.
Architecture: BERT.
Model training with limited computational resources (gradient accumulation, etc.).
Evaluation of the model performance in production.
Model training pipeline, helping other developers to learn using the pipeline.

Experiments

Argued for blind experimentation and got it adopted
Produced code for blinding and post-evaluation unblinding
Statistical significance testing using statistical criteria and numerical simulation (bootstrap), including cases of multiple comparisons
Sample size (statistical power) estimation with numerical simulation (bootstrap)
Manuals and training in statistics and use of the above code

Large language models (LLM): OpenAI GPT, Anthropic Claude, etc.

Evaluation and comparison of various LLMs (Llama, SeamlessM4T, Mistral and others)
Using LLMs for text translation and editing, translation quality assessment, and glossary compilation
Prompt engineering
Fine-tuning

Information retrieval

PDF parsing
Querying a corporate knowledge base
Retrieval augmented generation.

Machine translation quality assessment

Comparison of various MT quality metrics (hLEPOR, COMET, BERTScore)
Using LLM to assess translation quality

Corporate text-to-speech service

Access via a web-interface
Backend prototype

Machine interpretation (speech-to-speech translation)

Protype with a client-server architecture
Proof-of-concept testing

Other natural language processing tasks

Tokenizer comparison and selection for different tasks (NLTK, SpaCy, tiktoken, BPE, WordPiece)
Named entity recognition (NER)

Databases and analytics. A retail business within the same holding company.

Designed a schema for a sale and delivery database
Data loading
Regular analytics:

Marketing activity prioritization
On-shelf availability assessment

Back to top

Background

Higher education: Faculty of Physics, Moscow State University. Degree: diploma of higher education (equivalent of Master of Physics).

Data Science reskilling course. Tomsk State University.

Achievements:

I was the first to complete training out of ~160 students

Students learning faster presented their projects to other students to help them progress on the course. I presented four projects ouf of 16

As a top student, I also checked other students' graduation projects (normally done by staff)

Other training:

Data Engineer course in the Sapiens Academy (ELT/ETL, DWH, Greenplum, Airflow, Clickhouse, Superset).

Recommender Systems in Practice bootcapmp. Higher School of Economics/Magnit.

Uplift Modeling course at Open Data Science.

Basics of Statistics and A/B Testing.

Docker. Karpov.courses

Stepic courses. All courses were completed with distinction, ranking among 1 to 6% top students. Certificates.

Python Basics and Application
Programming in Python
Data Analysis in R
Basic Programming in R
Basics of Statistics, parts 1, 2 and 3
Intro to Data Science and Machine Learning
Interactive SQL Simulator
Intro to Linux
BI/Data Engineer. Fast Start

Back to top