Максим

Maksym Ivashchyshyn
Data Scientist
[відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | Ternopil, Ukraine

EDUCATION

Ternopil Ivan Puluj National Technical University – Bachelor's degree in Cybersecurity 2021 – 2025
Ternopil Ivan Puluj National Technical University – Master's degree in Cybersecurity 2025 – Present

PROFESSIONAL SKILLS
Programming & Data Analysis
● Python: Primary language for statistical analysis, data mining, and building advanced machine learning pipelines.
● Libraries: Pandas, NumPy, SciPy, Scikit-learn, Matplotlib, Seaborn (utilized for complex numerical
computations, data cleaning, exploratory data analysis (EDA), and visualization).
● Database & Data Management: PostgreSQL, BigQuery, Vector DBs (ChromaDB, Qdrant) for efficient data
retrieval and unstructured text processing.
Machine Learning & Statistical Modeling
● Regression & Forecasting: Implemented Linear Regression (OLS, Ridge, Lasso) and advanced ensemble
methods like XGBoost for predictive analysis and time-series forecasting (SARIMA, Amazon Chronos).
● Classification Models: Designed and optimized predictive models using Logistic Regression, SVM, Decision
Trees, XGBoost and K-NN, employing cross-validation and hyperparameter tuning to enhance performance.
Computer Vision
● Object Detection: Trained and fine-tuned YOLO models for multi-class object detection and classification tasks.
● Inference Optimization: Applied SAHI to improve detection accuracy on small objects in high-resolution
images.
● Experiment Tracking: Used ClearML for experiment tracking, model versioning, and metrics comparison across
training runs.
Deep Learning
● PyTorch, TensorFlow: Built and scalable deep learning models, utilizing flexible computational graphs for
custom model training and experimentation.
● Transformers: Applied Transformer-based models for time-series prediction and sequence modeling.
Natural Language Processing & Generative AI
● LLM Agents & RAG: Designed autonomous agent workflows and Retrieval-Augmented Generation systems
using Claude/GPT/Gemini for context-aware analysis.
● NLP Techniques: Utilized deep learning approaches for sentiment analysis and parsing text data.
● Tools: LangChain, smolagents, Vector Databases for semantic search.

PROFESSIONAL EXPERIENCE

Data Scientist, Amazinum – Ternopil, Ukraine Oct 2024 – Jan 2026
Loan Default Prediction System
● Built an end-to-end ML pipeline for credit risk assessment, integrating multi-source financial data and
transforming unstructured banker commentary into clean binary labels for model training.
● Engineered 10+ predictive features covering financial stability; handled missing data via contextual encoding and
similarity-based imputation.
● Trained an XGBoost classifier and addressed severe class imbalance (~90/10) via scale_pos_weight balancing;
tuned the decision threshold to optimize the precision/recall trade-off for the business use case.
● Deployed the model as a production-ready FastAPI service with real-time/batch prediction endpoints,
containerized with Docker, including an automated evaluation framework for performance monitoring.

Financial Agent Development
● Designed and implemented autonomous financial agents using Claude 4 Sonnet and the smolagents framework.
● Developed a hybrid forecasting engine integrating XGBoost and SARIMA models.
● Built a RAG architecture with ChromaDB to provide AI agents with context-aware access to financial data,
enhancing the accuracy of investment recommendations.
● Automated DCF valuation and portfolio construction using skfolio.
● Deployed ML models and AI services via a high-performance FastAPI backend, utilizing Docker for consistent
environment management and reproducibility.
Intelligent Document Parser
● Built an automated data processing pipeline for LLM fine-tuning, capable of generating structured JSONL
training pairs and cutting manual preparation efforts by 80%.
● Built a LLM validation system (Python, LangChain, Gemini) achieving 90%+ accuracy against ground truth data.
● Implemented automated confidence scoring with detailed reasoning generation. Engineered multi-criteria
evaluation prompts achieving high validation accuracy.
● Implemented a pipeline that maps AI reasoning chains directly to extracted data fields for precise explainability.
● Designed a robust prompt architecture for structured outputs.
Blueprint Object Detection & Classification
● Trained and fine-tuned YOLO models to detect and classify 50+ furniture and architectural element classes on
floor plan drawings.
● Built the data pipeline end-to-end: converted source PDF blueprints into a ~600–1,000 image dataset, coordinated
annotation quality with a dedicated data labeling/QA team, and handled data cleaning, augmentation, and class
balancing.
● Applied SAHI to boost detection accuracy on small objects within high-resolution blueprint images.
● Tracked experiments and model versions with ClearML, reaching approximately 0.85 mAP@50 and 0.65
mAP@50-95 on the validation set, with precision/recall around 0.85–0.90 and an F1-score of ~0.87.

AI/ML Engineer, SmartTender – Ukraine Jan 2026 – Present

Core AI Services Development
● Built a production LLM interface on Google Vertex AI/Gemini, with pre-flight token budgeting and structured
JSON schema validation for reliable, well-formed model outputs.
● Engineered a vector embedding pipeline for semantic product similarity search and deduplication across a large-
scale product catalog, with caching to speed up repeated similarity queries.
● Developed a price-weighted analysis (RDN) microservice for electricity tenders – extracting tender conditions
from BigQuery and computing weighted-average pricing and margin calculations.
● Implemented a geographic distance checker for gas station tender requirements, combining LLM-based
requirement parsing, geocoding, and road-distance filtering against real procurement data.
● Integrated MLFlow for distributed tracing and observability across AI operations.

Tender Item Classification
● Built a semantic search system in PostgreSQL that matches products to the most relevant category (out of 1000+
categories) by comparing AI-generated embeddings, instead of relying on exact keyword matches.
● Used the top search results as suggestions for the AI model, so instead of picking from thousands of categories, it
only had to choose the best match from a short list of likely candidates.Added caching for vector search results to
cut redundant embedding generation and lookups on repeat tenders.
● Engineered an adaptive batch classification system with multi-level fault tolerance for reliable classification at
scale.
● Designed the PostgreSQL/SQLAlchemy integration layer for storing and managing classification results across
the pipeline.