Work

Detecting AI-Generated Code

Comparing three method families ranging from traditional classifiers, to graph-based methods, to deep learning embedding-based classification.

Role

Research and Development Lead

Stack

Python / PyTorch / PyTorch-Geometric / Tree-Sitter / TensorBoard

• Traditional machine learning classifiers
• Syntax tree and graph-based representations
• Deep learning embedding-based classification

Universal Launcher

A Windows desktop launcher that centralizes applications, project folders, shortcuts, notes, reminders, and frequently used tools.

Role

Team Lead/Lead Dev

Stack

C# / .NET Framework 4.8 / Windows Forms

• Launches executables, shortcuts, internet shortcuts, folders, and nested applications from one library.
• Opens development projects through their parent IDE and supports organizing entries with favorites, categories, search, and recently used items.
• Includes notes, reminders, light and dark modes, and per-user serialized application data.

2D Autonomous Driving

A reinforcement-learning comparison of PPO, A2C, and QRDQN agents trained to navigate simulated highway and merge environments.

Role

Researcher and Developer

Stack

Python / PyTorch / Stable-Baselines3 / Highway-Env / TensorBoard / BoTorch

• PPO, A2C, and QRDQN agent training and evaluation
• Bayesian hyperparameter optimization
• Highway and merge environment comparison

Online Library

An ASP.NET Core library system for managing books, authors, categories, borrowing carts, borrowing histories, and member notifications.

Role

Co-Developer

Stack

C# / ASP.NET Core / Razor / Azure

• Organizes the application into domain, repository, service, and web layers.
• Supports catalog management, borrowing carts, borrowing and return histories, and notifications.
• Includes a separate administration application for members and borrowing records.

Country Collaboration Prediction

A graph machine-learning study using GDELT event data and temporal country-interaction networks to predict cooperative relationships between countries.

Role

Co-Researcher and Co-Developer

Stack

Python / Jupyter / NetworkX / Node2Vec / scikit-learn / TensorFlow

• Monthly country-interaction graphs derived from GDELT event records
• Node2Vec embeddings and similarity-based link features
• Classification experiments using linear, tree-based, support-vector, and neural-network models

Historic and Cultural Locations in Macedonia

A map-based web application for discovering historic and cultural locations in Macedonia, with category filtering, search, and directions.

Role

Lead Developer and Lead Designer

Stack

React / Java / Spring Boot / Leaflet / Docker / Azure

• Displays historic and cultural locations through an interactive Leaflet map.
• Supports category filtering, search queries, and routing from the user's location to a selected destination.
• Separates the React client, Spring Boot REST API, and proxy into containerized services.

From Syntax Trees to Embeddings: A Comparative Study of AI-Generated Code Detection

A comparative study of AI-generated code detection using engineered syntax-tree features, CodeBERT embeddings, and graph-based syntax representations.