Maria Afanasyeva

Data Science · Senior · Management: 1-2 years (1-5 people) · Lisbon, Portugal · UTC+00:00 · In office, Partially remote, Fully remote · Relocation · Work permits: Portugal, Russia · Citizenship: Russia · English: C1 · Can start in 2 weeks

QA Automation Engineer

9+ years exp

Management: 1-2 years (1-5 people)

5+ projects

Python

Java

Data Wrangling & Cleaning: 8/10

Exploratory Analysis & Visualization: 8/10

Predictive Modeling: 7/10

Active 5 days ago

Download CVCV

Overview

Technical skills

Projects

Timeline

Roles

Overview

An experienced ML practitioner (senior-level notebook work) focused on entity resolution and duplicate detection with strong feature engineering and evaluation skills. The single strongest proven skill is production-minded entity-resolution feature engineering and pairwise modeling as demonstrated in AI-customer-deduplication/EDA_and_model.ipynb (profile-level aggregation, fuzzy/email/phone normalization, leakage-aware pair generation, and graph-based clustering). There is limited evidence of production engineering around testing, CI/CD, scalable candidate-generation (blocking at scale), and model serving/monitoring in public artifacts.

Phone

Technical skills

Languages

Python

Java

DevOps

Terraform

Ansible

Git

Jenkins

Selenium

Pytest

Locust

AI/ML

Claude

Model Context Protocol

Pandas

NumPy

Scikit-learn

TensorFlow

CatBoost

Other

PostgreSQL

Matplotlib

LLM

Jupyter Notebook

Docker

GitLab CI

Linux

CI/CD

Machine Learning

NLP

Projects

AI StyleLink

Nov 2025 to Present 11 Months

Co-founder and AI Engineer

Developed an ML recommendation architecture for fashion personalization using LLM-based query generation, FashionCLIP retrieval, and compatibility filtering combined with ruBERT-based size prediction.

Collected and normalized a large set of customer reviews across multiple product groups.

Trained an NLP model to classify clothing size deviations based on the normalized review data.

Jupyter Notebook

Python

TensorFlow

Scikit-learn

Keras

Just Lives Studio (Game Development Studio)

Feb 2022 to Dec 2023 1 Year 10 Months

Founder and CEO

Developing iOS/Android mobile games - interactive stories.

I am the founder of this project, and also a developer (python - RenPy engine), designer and partially an artist. Also as a CEO I curated aspiring professionals working with me

Python

Ren'Py

GitHub Actions

Android SDK

Mail.ru Test Automation Course

Mar 2021 to May 2021 2 Months

Autotests and DevOps

Built a final capstone repository focused on automated testing and DevOps practices. Implemented an automation setup intended to run tests through a CI/CD-oriented workflow. Organized the project as a reusable starting point for automated execution and maintenance.

Python

Pytest

Docker

Docker Compose

Jenkins

MySQL

Flask

Timeline

Senior QA Automation Engineer (Mobile) · Senior

Nexters · Full-Time

Oct 2024 to Present 2 Years Limassol Fully remote

Built an agentic mobile test automation approach using Claude to interpret game design documents and validate recurring game events against requirements.

Implemented an MCP server to power an internal assistant for reading and updating configurations.

Improved CI/CD workflows and tooling UX to increase automation adoption, and developed an AI-assisted tool to analyze failed test reports and accelerate test maintenance.

Expanded mobile test accuracy and coverage while maintaining an in-house mobile device farm.

Claude

Model Context Protocol

Python

GitLab CI

Android SDK

Moscow Institute of Physics and Technology

Master's Degree · Artificial Intelligence

2025–2027 Moscow, Russia

Senior QA Automation Engineer · Senior

Karta · Full-Time

Sep 2022 to Oct 2024 2 Years 1 Month Denver Fully remote

Established test automation from scratch for fintech products, reaching high automated coverage and expanding validation across key integrations.

Scaled the automation process to improve release speed and reliability while increasing overall product quality.

Mentored manual QA engineers and improved documentation to speed up onboarding.

Designed and ran performance testing using Locust.

Locust

Python

GitHub Actions

Pytest

Selenium

Requests

Copilot

Allure Reporting

Moscow State University (MSU)

Master's Degree · Fundamental Economics

2020–2022 Moscow, Russia

Senior Software Development Engineer in Test · Senior

SBER · Full-Time

Nov 2020 to Mar 2022 1 Year 4 Months Moscow Partially remote

Created web regression automation covering UI, API, and database layers using Python-based testing and browser automation.

Implemented automated execution and reporting using Jenkins pipelines, enabling QA teams to focus on new features.

Restructured the test framework to keep pace with rapid product changes.

Owned final quality for a nationwide census-related rewards project, adapting testing priorities immediately before launch.

Python

Pytest

Selenium

Jenkins

Jira

Java

Terraform

Requests

Middle Maintenance Engineer (DevOps) · Middle

SBER · Full-Time

Nov 2019 to Oct 2020 11 Months Moscow In office

Automated deployment of hundreds of virtual machines using Terraform and Jenkins pipelines.

Developed a custom internal Terraform provider in Go and expanded infrastructure capabilities.

Automated Linux provisioning, configuration, and software installation with Ansible and Jenkins.

Added Java-based UI/API testing and integrated smoke tests into CI/CD workflows.

Terraform

Jenkins

Ansible

Python

Linux

Red Hat

Intern (DevOps) · Junior

SBER · Internship

Jul 2019 to Oct 2019 3 Months Moscow In office

Developed Java UI and API test automation for an internal employee-offers platform.

Configured Jenkins pipelines to enable automated test execution.

Contributed to integrating automated runs into the team’s CI workflow for faster feedback.

Jenkins

Python

All-Russian Academy of Foreign Trade

Bachelor's Degree · International Finance

2014–2018 Moscow, Russia

Intern QA Automation · Junior

CAE Fidesys · Internship

In office

Migrated API tests from JavaScript to Python and increased automated coverage. Improved reliability of HTTP-based validations by expanding the existing request testing logic. Supported ongoing automation development to strengthen regression quality.

Python

QA Automation Engineer · Middle

Vertical Takeoff · Full-Time

In office

Built UI test automation from scratch for an educational CRM using Python and Selenium. Focused on creating maintainable browser-based automated checks to support ongoing product quality. Worked on end-to-end test flow stabilization for reliable regression runs.

Python

Selenium

Senior Data Scientist Confidence: High ML Practitioner

Statistical Rigor

5/10

Correct use of statistics

Moderate statistical rigor: uses AUC and precision/recall thresholding, reports clustering quality (ARI/NMI) and inspects feature distributions, but lacks formal uncertainty quantification, hypothesis testing with assumptions, or robust cross-validation beyond a single train/val/test split.

Evidence

AI-customer-deduplication/EDA_and_model.ipynb: fuzzy_auc AUC computation and fuzzy_auc DataFrame

AI-customer-deduplication/EDA_and_model.ipynb: threshold_df and threshold-based precision/recall selection

Data Wrangling & Cleaning

8/10

Preparing and cleaning data

Strong data wrangling and cleaning: careful profile-level aggregation, robust missing-value checks, normalization helpers, phone/email parsing, JSON parsing for realtime fields, and token merging for list-valued features.

Evidence

AI-customer-deduplication/EDA_and_model.ipynb: profile-level aggregation block building profile_df using first_non_empty/merge_unique_lists

AI-customer-deduplication/EDA_and_model.ipynb: normalize_phone, email_local_part, parse_realtime functions

Exploratory Analysis & Visualization

8/10

Exploring and visualizing data

High-quality exploratory analysis and visualization with clear written findings after sections; EDA is hypothesis-driven and ties findings to design choices for features and modeling.

Evidence

AI-customer-deduplication/EDA_and_model.ipynb: 'Findings' text after missing-value, email, phone and cluster-size analyses

AI-customer-deduplication/EDA_and_model.ipynb: plots and completeness_summary showing duplicate rates by completeness

Predictive Modeling

7/10

Building models that predict

Solid predictive modeling work: careful pair generation with leakage control (entity-level split), baseline decision tree and CatBoost experiments, feature-importance analysis, thresholding strategy and graph-based clustering of predicted matches.

Evidence

AI-customer-deduplication/EDA_and_model.ipynb: train/val/test split by entity_id to prevent leakage (train_test_split on entities)

AI-customer-deduplication/EDA_and_model.ipynb: model_boost (CatBoostClassifier) training and feature_importance, plus DecisionTree baseline

Business Insight & Impact

7/10

Turning analysis into business value

Business-aware analysis: problem is well framed for marketplace data, error-cost reasoning prioritizes precision vs recall, and operational decisions (auto-merge vs manual review) are derived from validation metrics.

Evidence

AI-customer-deduplication/README.md: problem framing for marketplace customer profiles and intended business impact

AI-customer-deduplication/EDA_and_model.ipynb: explicit threshold selection and narrative prioritizing precision (auto_merge and manual_review thresholds)

Reproducibility & Notebook Hygiene

5/10

Clean, repeatable analysis

Moderate reproducibility: random seeds and requirements.txt are present and deterministic choices are used, but workflows remain notebook-centric with hard-coded paths and limited packaging, testing, or data/versioning infrastructure.

Evidence

AI-customer-deduplication/requirements.txt: pinned core dependencies list

AI-customer-deduplication/EDA_and_model.ipynb: use of RANDOM_STATE/seed and deterministic pair generation functions

Industries

Commerce· Middle

Technologies

CatBoost

Jupyter Notebook

Scikit-learn

Matplotlib

Pandas

NumPy

Recommendations

Develop a production-ready, blocked candidate-generation pipeline (e.g., scalable blocking/indexing) and integrate it with the existing pairwise feature builder to handle much larger profile volumes.
Implement a repeatable training pipeline: move notebook code into modular Python packages, add unit tests for key functions (normalizers, pair generation), and create CI to run model training and evaluation.
Add monitoring and post-merge quality checks: build labeling feedback loops, drift detection on key similarity features, and a conservative rollback/manual-review UI for merges.
Integrate a feature-store or persistent feature artifacts (Feast or similar) to enable consistent feature computation at training and serving time.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

AI-customer-deduplication waste-classification-model

Middle AI/ML Engineer Confidence: Medium ML Engineer

Customer-data ML engineer at a middle level specializing in end-to-end entity resolution and duplicate-profile detection. The strongest proven skill is practical feature engineering and pairwise comparison pipelines demonstrated by the profile-level aggregation, normalization functions and pairwise feature construction in AI-customer-deduplication/EDA_and_model.ipynb. There is no public evidence of production model serving, MLOps, distributed training, or custom neural architectures.

Model Architecture & Training

4/10

How well models are designed and trained

Applied model development and training with sensible choices (DecisionTree and CatBoost, class weighting, eval_set and threshold selection) but no custom architectures or distributed training.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: DecisionTreeClassifier training cell am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: CatBoostClassifier(...) fit call and eval_set usage

Data Pipeline & Feature Engineering

6/10

How data is prepared for models

Strong, practical data pipeline and feature engineering for entity resolution including profile-level aggregation, normalization, token extraction and pairwise feature construction.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: profile_df aggregation using first_non_empty, merge_unique_lists and profile-level feature creation am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: normalize_text, normalize_phone, parse_realtime and extract_token_value functions am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: pairwise construction and build_features / pair_features feature engineering cells

Experimentation & Evaluation

5/10

How results are measured and tested

Thoughtful experimentation and evaluation with entity-level splits to avoid leakage, AUC/ROC and threshold sweep, and cluster-quality metrics for downstream effect assessment.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: train/val/test split by entity_id am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: auc calculation, roc_auc_score usage and threshold sweep producing threshold_df am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: graph clustering evaluation with adjusted_rand_score and normalized_mutual_info_score

MLOps & Deployment

1/10

How models are shipped to production

Minimal MLOps or deployment artifacts; clustering and model selection are presented as offline notebooks without serving, versioning, or monitoring components.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: networkx-based graph clustering to produce predicted_entities (offline pipeline)

Computational Efficiency

1/10

How efficiently computing resources are used

Little to no evidence of computational-efficiency engineering such as GPU optimization, batching, quantization or profiling; a pure-Python Levenshtein fallback exists but no performance benchmarking.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: levenshtein_distance_py fallback implementation and conditional rapidfuzz usage

Research Depth & Innovation

2/10

Depth of research and new ideas

Solid applied engineering and domain-aware feature design but no novel algorithmic contributions or paper-level reproductions.

Evidence

am42saturday/AI-customer-deduplication/EDA_and_model.ipynb: detailed feature_catalog and discussion of feature priorities and business-safe thresholding

Expertise

Graph Machine Learning (GNN)· Middle

Technologies

Claude· since 2024

Terraform· since 2019

Ansible· since 2019

Model Context Protocol· since 2024

NLP

GitLab CI

CI/CD

Jenkins· since 2019

Git· since 2021

Docker

LLM

Linux

Machine Learning

Recommendations

Develop production-grade inference and MLOps for entity resolution including model serving, versioning, and drift monitoring
Build scalable blocking and candidate-generation modules to reduce pairwise comparisons and enable large-scale deployment
Extend the graph-based pipeline with graph-ML methods or GNN prototypes for end-to-end learned clustering where appropriate
Add experiment tracking and reproducible pipelines (W&B or MLflow) and integrate dataset/version control for robust evaluations

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

AI-customer-deduplication spark_hw5