Yaroslav Matushenko

Intern Data Scientist (SpeechTech)

X5 Tech

Technology division of Russian retailer X5 Group.

Built speech recognition evaluation and fine-tuning workflows, contributed to an internal speech platform pilot, and developed a synthetic-data keyword spotting MVP.

• Designed a benchmark for speech recognition systems, including LLM-based evaluation, enabling more granular assessment of model performance across application domains. • Fine-tuned open-source speech recognition models for domain-specific use cases in both offline and streaming settings, enabling integration into business workflows. • Contributed to automating the team’s technical workflows, including the end-to-end pipeline for model training, evaluation, and production deployment.

ML Engineer (Medical CV)

MRTAI

Developed MRI segmentation and stroke-localization models, built a deployed service and data pipeline, and supported clinical evaluation through PACS integration.

• Researched and developed methods for analyzing 2D/3D multimodal brain MRI for automated lesion segmentation. • Developed and trained multiclass and binary MRI segmentation models based on nnU-Net; conducted experiments with architectures, preprocessing, augmentations, and training parameters. • Designed an end-to-end medical data preparation pipeline: DICOM → NIfTI conversion, dataset quality control and structuring, and preparation of training and validation data.

Data Scientist

HeartScan

Built SCG segmentation and classification models and an end-to-end event-detection pipeline, with reported 93–99% metrics and dataset growth from 25 to 1,500 hours.

• Developed a model for segmenting aortic valve opening events in seismocardiograms (SCG); Dice = 96%, IoU = 93%. • Trained a TS2Vec-based model for noisy SCG classification; achieved Accuracy = 95% and expanded the representative dataset from 25 to 1,500 hours. • Migrated model APIs and inference from Flask to FastAPI and NVIDIA Triton Inference Server. • Implemented an end-to-end ML pipeline for detecting aortic valve opening (AO) and closing (AC) events in SCG; AccuracyAOevent = 99%; Accuracyevent = 95% AC.

Researcher (NLP & CV)

Laboratory of Artificial Intelligence (SUAI)

Developed NLP pipelines for extracting entities, classifying regulatory-document sections, and searching a document database with RAG.

• Designed an information extraction pipeline for unstructured documents (regulations, instructions, and contracts), including cleaning, tokenization, and section splitting. • Implemented NER for extracting key entities such as job titles, departments, deadlines, and document numbers. • Built a RuBERT/RuRoBERTa-based classifier for categorizing pages and document fragments by type (introduction, requirements, responsibilities, deadlines). • Developed a RAG-based approach for searching a regulatory document database using a vector database.

B.Sc. in Biotechnical Systems and Technologie

SUAI