Evgeniy is available for hire
Hire EvgeniyEvgeniy Salskiy
Verified Expert in Engineering
Machine Learning Engineer and Developer
Yerevan, Armenia
Toptal member since April 6, 2026
Bio
Evgeniy is a senior machine learning engineer with 12 years of experience shipping production-ready systems. At Yandex, he guided a team of 10 to develop a real-time content classification platform processing millions of events per day and integrated LLMs into production, reducing annotation costs by 97%. Previously, Evgeniy also built anti-fraud and credit scoring systems in fintech from scratch, handling real-time inference, feature engineering on financial streams, and model monitoring.
Portfolio
Yandex
Python, PyTorch, Hugging Face, Transformers, BERT, Large Language Models (LLMs)...
ClearScale
Python, PyTorch, Scikit-learn, Amazon Web Services (AWS), Amazon SageMaker...
Microfinancial Organization Agora
XGBoost, CatBoost, Python, Credit Scoring, Fraud Detection, Anomaly Detection...
Experience
- Fraud Detection - 10 years
- Risk Modeling - 8 years
- Natural Language Processing (NLP) - 8 years
- Recommendation Systems - 8 years
- Large Language Models (LLMs) - 7 years
- Quantitative Trading - 6 years
- Time Series Forecasting - 6 years
- LangChain - 6 years
Preferred Environment
Machine Learning, Natural Language Processing (NLP), Trading, Energy Modeling, Fraud Prevention, Fraud Detection, Credit Risk, Risk Models, Amazon Web Services (AWS), Data Engineering
The most amazing...
...challenge I've tackled was integrating LLMs into a production system handling millions of events daily, achieving a 97% reduction in annotation costs.
Work Experience
Senior ML Engineer | Team Lead
2022 - 2026
Yandex
- Guided a team of 10 engineers in building an NLP-powered content classification system that processes 5+ million text documents daily with sub-100ms inference latency.
- Designed and deployed an LLM-based annotation pipeline using YandexGPT, replacing manual text labeling and reducing annotation costs by 97% while maintaining over 95% agreement with human annotators.
- Built a multistage NLP pipeline combining text classification, named entity recognition, and semantic similarity models—reducing unwanted content in search results by 78%.
- Deployed a reinforcement learning from human feedback (RLHF) workflow for LLM fine-tuning. Designed preference data collection, trained reward models, and optimized policy using proximal policy optimization (PPO) across 500,000+ labeled text pairs.
- Developed a comprehensive LLM evaluation framework measuring over 12 quality metrics, including fluency, factuality, relevance, and toxicity, across more than 50,000 test cases, enabling data-driven model iteration cycles.
Technologies: Python, PyTorch, Hugging Face, Transformers, BERT, Large Language Models (LLMs), Reinforcement Learning from Human Feedback (RLHF), Natural Language Processing (NLP), Prompt Engineering, Text Classification, Sentiment Analysis, Tokenization, Embeddings from Language Models (ELMo), Fine-tuning, Apache Kafka, A/B Testing, Model Evaluation, Energy Modeling, Vector Databases, LangChain, Ranking Models, Collaborative Filtering, Recommendation Systems, Data Governance, Data Management, Risk Management, Leadership, Agentic AI, Anthropic, Claude, Model Context Protocol (MCP), Qdrant, ColBERT, RAG Systems, Generative Artificial Intelligence (GenAI), OpenAI, APIs, Agentic RAG Systems, LangGraph, RAG Pipelines, FastAPI, Apache Spark, Elasticsearch, Git, GraphQL, Linux, Google Cloud Platform (GCP), Hyperparameter Tuning, API Integration, Benchmarking, Decision Modeling, Retrieval-augmented Generation (RAG)
Senior Machine Learning Engineer
2021 - 2022
ClearScale
- Delivered eight end-to-end machine learning projects across finance, healthcare, logistics, and retail, managing each from business requirements through production deployment.
- Built a real-time recommendation engine for an eCommerce client processing 2+ million daily user events, increasing click-through rate by 23% over a rule-based baseline.
- Developed a time series demand forecasting model for a logistics client, reducing inventory overstock costs by 15% across 200+ warehouse SKUs.
- Streamlined a machine learning project delivery process by creating standardized templates for requirements gathering and evaluation, reducing average delivery time by 15%.
- Implemented automated model performance reporting framework across all eight projects, enabling data-driven decision-making and reducing manual reporting by 80%.
Technologies: Python, PyTorch, Scikit-learn, Amazon Web Services (AWS), Amazon SageMaker, AWS Lambda, Docker, Recommendation Systems, Natural Language Processing (NLP), Time Series Forecasting, Energy Modeling, Vector Databases, Transactions, Content-Based Filtering, Data Governance, Data Management, Risk Management, California Consumer Privacy Act (CCPA), Cloud Governance, Collibra, OpenAI, APIs, FastAPI, Apache Spark, Elasticsearch, Git, GraphQL, Linux, Hyperparameter Tuning, API Integration, Benchmarking, Decision Modeling
Senior Data Scientist
2018 - 2021
Microfinancial Organization Agora
- Built a real-time anti-fraud scoring engine from scratch on streaming financial transactions, reducing fraud rate by 67% within the first six months of deployment.
- Developed a machine learning-based credit scoring model on financial time series data, reducing the loan default rate from 32% to 26% across a portfolio of more than 50,000 borrowers.
- Designed a feature engineering pipeline that extracts over 40 behavioral signals from transaction sequences, including velocity checks, device fingerprinting, and session patterns.
- Implemented a model monitoring system tracking Gini, ROC-AUC, and PSI metrics daily, with automated alerts triggered when model performance degrades beyond a 5% threshold.
- Spearheaded a team of three engineers across the full machine learning lifecycle, including data collection, feature engineering, model training, A/B testing, production deployment, and retraining.
Technologies: XGBoost, CatBoost, Python, Credit Scoring, Fraud Detection, Anomaly Detection, Time Series Analysis, Risk Modeling, Docker, Apache Kafka, Energy Modeling, Retrieval-augmented Generation (RAG), LangChain, Ranking Models, Content-Based Filtering, Data Management, Data Quality Governance, Risk Management, Leadership, Cloud Governance, APIs, FastAPI, Git, Linux, Hyperparameter Tuning, API Integration, Benchmarking, Decision Modeling
Data Scientist
2018 - 2018
Siberian Generating Company
- Developed short-horizon trading strategy models for the day-ahead energy market, optimizing bid placement across 24 hourly price slots.
- Built a time series forecasting pipeline for energy demand and price dynamics, reducing prediction error by 18% vs the previous Excel-based approach.
- Automated a daily reporting workflow that previously required three hours of manual Excel work, reducing report generation time to under 10 minutes.
- Analyzed historical price and load data across more than 12 months to identify seasonal patterns, peak demand drivers, and price anomalies.
- Delivered data-driven pricing recommendations that improved day-ahead market bidding accuracy, contributing to an estimated 8% revenue uplift during the test period.
Technologies: Python 3, Python, Pandas, NumPy, Scikit-learn, SQL, Time Series Analysis, Statistical Modeling, Excel VBA, Data Visualization, Machine Learning, Artificial Intelligence (AI), Data Engineering, Vector Databases, Transactions, Collaborative Filtering, Data Governance, Data Quality Governance, Risk Management, APIs, Git, Linux, Hyperparameter Tuning, Quantitative Trading, API Integration
CV Engineer
2015 - 2018
Spectr
- Developed a text recognition system for medical documents, including diagnoses and analyses, achieving over 95% accuracy on noisy handwritten inputs.
- Built an OCR pipeline for identity documents, including passports, birth certificates, and insurance policies, with automated field extraction and validation.
- Designed a document recognition system for construction estimates and engineering drawings, parsing complex table structures into structured data.
- Implemented a preprocessing pipeline for scanned documents, including noise reduction, binarization, layout analysis, and text line segmentation.
- Delivered production-ready recognition systems across 5+ document types, each requiring domain-specific training data and post-processing rules.
Technologies: Python 3, OpenCV, Tesseract, Optical Character Recognition (OCR), Computer Vision, Image Processing, Natural Language Processing (NLP), Text Recognition, Collaborative Filtering, Git, Linux, FastAPI, API Integration
Experience
Real-time Face Swap Engine
https://github.com/UltraEvgeny/webcam_face_replaceI built a real-time face replacement system for webcam video streams using deep learning. I implemented face detection with MTCNN, facial landmark extraction with dlib, and face swapping using a custom autoencoder architecture. The pipeline processes 30 FPS video with sub-50ms latency per frame.
I also handled edge cases, including partial occlusion, varying lighting conditions, and head rotation up to 45 degrees. I optimized inference for consumer GPUs using TensorRT, reducing model size by 60% while maintaining visual quality. Finally, I built the solution as a standalone Python application with OpenCV for video capture and rendering.
I also handled edge cases, including partial occlusion, varying lighting conditions, and head rotation up to 45 degrees. I optimized inference for consumer GPUs using TensorRT, reducing model size by 60% while maintaining visual quality. Finally, I built the solution as a standalone Python application with OpenCV for video capture and rendering.
Solana DEX Arbitrage Backtesting Framework
https://github.com/UltraEvgeny/sol-arb-backtestI designed and implemented a backtesting framework to evaluate arbitrage opportunities across Solana DEX protocols, including Raydium, Orca, and Jupiter. The system processes historical on-chain transaction data to identify cross-pool price discrepancies, simulates trade execution with realistic gas costs, slippage, and MEV competition, and calculates risk-adjusted returns.
I built a vectorized simulation engine using NumPy and pandas, processing 10+ million transactions per backtest run. I also implemented statistical analysis of arbitrage window duration, profitability distribution, and optimal position sizing. Results showed viable arbitrage windows averaging 200ms with a median profit of 0.3% per trade before costs.
I built a vectorized simulation engine using NumPy and pandas, processing 10+ million transactions per backtest run. I also implemented statistical analysis of arbitrage window duration, profitability distribution, and optimal position sizing. Results showed viable arbitrage windows averaging 200ms with a median profit of 0.3% per trade before costs.
ML-based CAPTCHA Recognition System
I developed a deep learning pipeline for automated CAPTCHA recognition, achieving 94% accuracy on distorted alphanumeric CAPTCHAs. I built a custom CNN architecture with an attention mechanism for character-level segmentation and recognition.
The training pipeline encompassed synthetic CAPTCHA generation with configurable distortion parameters, including rotation, noise, overlapping characters, and color variation, producing more than 500,000 training samples. I implemented data augmentation strategies that improved accuracy by 12% over the baseline. I deployed the solution as a REST API service with FastAPI, handling more than 100 requests per second and achieving an average inference time of 15ms per image.
The training pipeline encompassed synthetic CAPTCHA generation with configurable distortion parameters, including rotation, noise, overlapping characters, and color variation, producing more than 500,000 training samples. I implemented data augmentation strategies that improved accuracy by 12% over the baseline. I deployed the solution as a REST API service with FastAPI, handling more than 100 requests per second and achieving an average inference time of 15ms per image.
Facial Verification System for Financial Identity Scoring
I built a facial verification pipeline integrated into a credit scoring workflow for a microfinance organization. The system compares applicant selfie photos against ID document photos to detect identity fraud.
I implemented face detection using RetinaFace, feature extraction with ArcFace embeddings (512-dimensional vectors), and cosine distance for similarity scoring. I achieved a 99.2% true positive rate at a 0.1% false positive rate on a test set of 15,000 verification pairs. I also added a liveness detection module to prevent photo-of-the-photo spoofing attacks, reducing identity fraud attempts by 45%. The solution was deployed to AWS Lambda for serverless inference, achieving a cold start time of under three seconds.
I implemented face detection using RetinaFace, feature extraction with ArcFace embeddings (512-dimensional vectors), and cosine distance for similarity scoring. I achieved a 99.2% true positive rate at a 0.1% false positive rate on a test set of 15,000 verification pairs. I also added a liveness detection module to prevent photo-of-the-photo spoofing attacks, reducing identity fraud attempts by 45%. The solution was deployed to AWS Lambda for serverless inference, achieving a cold start time of under three seconds.
Production LLM Quality Evaluation Framework
I designed and built a comprehensive evaluation framework to measure LLM output quality in production content classification pipelines. The system tracks 12+ quality metrics, including factual accuracy, relevance, fluency, toxicity, and task-specific rubrics across 50,000+ test cases.
I implemented automated regression testing that catches quality degradation within 24 hours of model or prompt changes. I also built an A/B testing infrastructure to compare model versions with statistical significance testing (p<0.05). The framework reduced manual quality review time by 80% and enabled data-driven prompt optimization cycles, improving the classification F1-score from 0.82 to 0.91 over six months.
I implemented automated regression testing that catches quality degradation within 24 hours of model or prompt changes. I also built an A/B testing infrastructure to compare model versions with statistical significance testing (p<0.05). The framework reduced manual quality review time by 80% and enabled data-driven prompt optimization cycles, improving the classification F1-score from 0.82 to 0.91 over six months.
Education
2007 - 2014
Master's Degree in Mathematics and Computer Science
Kemerovo State University - Kemerovo, Russia
Certifications
SEPTEMBER 2018 - PRESENT
Professional Certificate in Computer Science
MIPT/Yandex/Coursera
Skills
Libraries/APIs
OpenCV, Pandas, NumPy, Scikit-learn, XGBoost, CatBoost, PyTorch
Tools
Amazon SageMaker, Collibra, Claude, Git
Languages
Python 3, Python, SQL, Excel VBA, GraphQL
Frameworks
MediaPipe, LangGraph, Apache Spark
Paradigms
Anomaly Detection, Real-time Systems, Model Context Protocol (MCP), Rule-based Programming
Platforms
Amazon Web Services (AWS), Docker, Apache Kafka, AWS Lambda, Google Cloud Platform (GCP), Vertex AI, Linux
Storage
PostgreSQL, Elasticsearch
Other
Machine Learning, Natural Language Processing (NLP), Trading, Energy Modeling, Fraud Prevention, Fraud Detection, Credit Risk, Risk Models, Data Engineering, Tesseract, Optical Character Recognition (OCR), Computer Vision, Image Processing, Text Recognition, Time Series Analysis, Statistical Modeling, Data Visualization, Artificial Intelligence (AI), Credit Scoring, Risk Modeling, Recommendation Systems, Time Series Forecasting, Hugging Face, Transformers, BERT, Large Language Models (LLMs), Reinforcement Learning from Human Feedback (RLHF), Prompt Engineering, Text Classification, Sentiment Analysis, Tokenization, Embeddings from Language Models (ELMo), Fine-tuning, A/B Testing, Model Evaluation, Mathematics, Differential Equations, NVIDIA TensorRT, Backtesting, Statistical Analysis, Decentralized Finance (DeFi), Solana, Arbitrage, Market Microstructures, Convolutional Neural Networks (CNNs), Data Augmentation, Deep Learning, Facial Recognition, Collaborative Filtering, Content-Based Filtering, Ranking Models, Transactions, LangChain, Vector Databases, Retrieval-augmented Generation (RAG), Quantitative Trading, Data Governance, Data Management, Data Quality Governance, Risk Management, Leadership, California Consumer Privacy Act (CCPA), Cloud Governance, Agentic AI, Anthropic, Qdrant, ColBERT, RAG Systems, Generative Artificial Intelligence (GenAI), OpenAI, APIs, API Integration, Benchmarking, Hyperparameter Tuning, OpenAI GPT-4 API, Decision Modeling, Agentic RAG Systems, RAG Pipelines, FastAPI
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring