Filin

Nazarii Filin
Machine Learning & Software Engineer
Klagenfurt, Austria | [відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація»)
Links: GitHub

TECHNICAL SKILLS

Languages: Python, C++, SQL
Machine Learning & AI: PyTorch, Hugging Face Transformers, TensorFlow (basic), Torchvision, Scikit-learn,
Pandas, Matplotlib, Joblib
ML Concepts: Deep Learning, Transformer Architecture, Multi-Head Attention, Bi-Encoder/Cross-Encoder Retrieval,
Contrastive Learning (InfoNCE), Dense Retrieval, Vector Search, CNNs
Backend & Databases: FastAPI, SQLAlchemy 2.0, PostgreSQL, SQLite, Qdrant
Systems, DevOps & Security: Docker, Docker Compose, Git, GitHub Actions, CI/CD, SARIF, SAST, AST Parsing
Software Engineering Fundamentals: OOP, SOLID, CMake, P2P Networking, Backend Architecture, Linear Algebra
Spoken Languages: English (B2, IELTS 6.0), Ukrainian (C2, native), Russian (C2, native)

EDUCATION

University of Klagenfurt Klagenfurt, Austria
B.Sc. in Artificial Intelligence and Robotics March 2025 - Present

SELECTED PROJECTS

Custom Neural RAG Pipeline - semantic search and text retrieval platform
Python, PyTorch, Hugging Face Transformers, Qdrant, FastAPI, Docker, Docker Compose

Implemented a custom Bi-Encoder Transformer in PyTorch with Pre-LayerNorm, Multi-Head Attention, Feed-
Forward layers, and masked mean pooling for dense text retrieval.
Fine-tuned the retrieval model on 50,000 CodeSearchNet samples using InfoNCE contrastive loss, improving
Recall@1 from 11.5% to 65.9% and Recall@10 from 27.1% to 87.8% on a held-out test set.
Engineered a two-stage retrieval pipeline combining Qdrant ANN search for Top-50 candidate retrieval with Cross-
Encoder reranking to select the Top-5 most relevant documents before generation.
Integrated context-grounded answer generation using an external LLM through the Hugging Face Hub SDK, with
prompt-based abstention when the retrieved context is insufficient.
Implemented pretrained weight transfer from Hugging Face MiniLM into the custom Transformer architecture.
Containerized the end-to-end ML pipeline providing reproducible local deployment and REST API access.

Surface Defect Classifier - computer vision anomaly detection
Python, PyTorch, Torchvision, FastAPI, SQLAlchemy 2.0, PostgreSQL, Docker, Scikit-learn

Developed a custom Convolutional Neural Network (CNN) in PyTorch to classify electron microscope images of
semiconductor wafers into 6 distinct surface anomaly classes.
Designed the model architecture using convolutional and max-pooling layers, optimizing with Adam and Cross-
Entropy loss, achieving 84% overall accuracy.
Evaluated model performance via confusion matrices and classification reports using Scikit-learn, supported by
an automated Torchvision data preprocessing pipeline.
Engineered a real-time REST API using FastAPI to serve the trained PyTorch model, enabling instant defect
classification and confidence scoring via HTTP requests.
Integrated SQLAlchemy 2.0 and a PostgreSQL database to persistently log prediction metadata and inspection
history.
Containerized the entire multi-service architecture (API and Database) using Docker and Docker Compose.

Nazarii Filin CV Page 1 of 2

Decentralized P2P Messenger - distributed communication network
C++, CMake, OpenDHT (Backend Network Engineer in a 3-person team)

Architected the core Kademlia-based P2P network layer using OpenDHT, fully decoupling the backend routing
logic from the UI and cryptography modules.
Developed the core backend networking module in C++, architecting the P2PNode class to control the node
lifecycle and establish decentralized connections.
Engineered network bootstrapping and peer discovery mechanisms utilizing OpenDHT to ensure reliable
connectivity across the distributed network.
Implemented robust user publication logic, including a background loop to maintain persistent profile visibility and
dynamically map UserIDs to network endpoints.
Designed the core packet transmission layer, managing direct message routing and secure inbox delivery.
Recognized by the course professor for developing a technically ambitious DHT-based prototype.

CryptoGuard SASTEngine - static application security testing
Python, AST, CI/CD, SARIF, GitHub Actions

Built a custom Static Application Security Testing (SAST) framework using Python's ast module to enforce secure
coding practices.
Developed context-aware heuristic rules to detect legacy cryptography (e.g., MD5) and insecure randomness,
reducing false positives.
Engineered a CI/CD security gate via GitHub Actions, automatically generating SARIF reports and blocking
vulnerable pull requests.

Nazarii Filin CV Page 2 of 2