Особисті дані приховані

Mykhailo Melnyk
ML engineer
Lviv, Ukraine | [відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | LinkedIn | GitHub

ML/AI Engineer with roots in academic machine learning: computer vision, predictive modeling – now applied to designing,
building, and deploying end-to-end ML systems across scientific computing, credit risk modeling, and LLM retrieval systems.
Currently focused on RAG/agentic systems. Open to ML/AI engineering roles.
Projects
Hierarchy-Guided RAG system
LangChain, LangGraph, Ragas, ZenML, MLflow, MongoDB, Qdrant June 2026 – present
● Designed a RAG system leveraging the notes’ natural hierarchical structure to guide retrieval strategy per query: low-
level chunks, high-level structure or bypassing retrieval entirely for general queries – improving context quality
● Orchestrated retrieval with LangGraph, combining query rewriting, cross-encoder reranking, and chunk location
metadata to surface more precise, better-grounded context
● Added an index-free agentic strategy that lets the agent freely traverse the hierarchy, receiving each time either the size-
bounded structure of descendants' titles or the leaf node contents themselves
● Structured the pipelines around FTI architecture with hash-based CDC for ETL and feature pipelines; built extendable,
versioned configs as the single source of truth across different retrieval strategies; used ZenML for pipeline observability
and MLflow for inference tracing and eval tracking
● Designed a custom LLM-as-judge answer-correctness metric (NLI entailment against atomic facts); improved answer
correctness from a 0.50 naive-chunking baseline to 0.737 (hierarchical RAG) and 0.755 (agentic) on an early eval set
Credit risk modelling system
Python, scikit-learn, optuna, FastAPI, Django, Docker Compose, LightGBM, XGBoost April – June 2026
● Built and deployed an end-to-end credit risk modelling system maximizing portfolio return through threshold-optimized
loan approval decision-making, achieving 1.73x aggregated return improvement on holdout set
● Deployed as two decoupled services – a FastAPI inference service and a Django web application with a database of
borrower information and loan decision tracking – containerized and orchestrated with Docker Compose
● Engineered feature-specific preprocessing across 151 raw variables, tailoring missingness, outlier, and encoding
strategies to each feature type and model family; found regression models to outperform classification formulation
Logo detection system | Python, OpenCV, Ultralytics, Roboflow, Albumentations, Google Colab May 2024
● Fine-tuned YOLOv8 on a custom Roboflow-annotated dataset with Albumentations augmentation using Google Colab,
achieved 0.991 precision and 0.983 recall; deployed locally for real-time video inference with interactive frame saving
Experience
Balloon path prediction application
Python, xarray, Qt, Leaflet.js, SQLite September 2024 – October 2025
● Owned the full lifecycle of a novel scientific computing application – requirements, architecture, algorithm design,
tooling, and end-to-end implementation – through independent research and iterative feedback with the commissioner
● Increased prediction accuracy by integrating multiple atmospheric data sources, flexible vertical path configuration and
model selection; ensured operational offline reliability via in-app forecast preload and priority-based failover system
● Implemented a full experimentation suite: launch and source parameters configuration, multi-path comparison on an
interactive map (including real flight paths), project saving and management
● Built a template-based parser for heterogeneous NetCDF/GRIB meteorological formats, enabling the system to ingest
new data sources without hardcoded format-specific logic
● Validated predictions against real balloon launches, presenting findings at SoftTech-2025
Skills & Certifications
LLM Engineering: LangChain, LangGraph, RAG pipeline design, Ragas, Qdrant
MLOps: ZenML, MLflow, Docker, Docker Compose
Development & Infrastructure: Python, SQL, JavaScript, FastAPI, Django, MongoDB, Git, Qt
Machine Learning & Data Science: scikit-learn, pandas, numpy, xarray, matplotlib, plotly, XGBoost, LightGBM, Optuna
Languages: English (professional working proficiency), Ukrainian (native)
Courses: Natural Language Processing with Classification and Vector Spaces (DeepLearning.AI), Financial Markets —
Robert Shiller, Yale University (Coursera)
Education
Kyiv Polytechnic Institute | Bachelor’s degree Kyiv, Ukraine | September 2021 – June 2025
Information systems software engineering
Relevant coursework: Neural Network Systems, Introduction to Data Science, Models and Methods of Artificial
Intelligence, Information Systems Infrastructure (AWS, Azure, VMware labs)