Worked od software for sales teams that automatically records calls, generates transcripts and summaries, and extracts key insights using AI. The system is integrated with CRM platforms and online meeting tools like Google Meet and Microsoft Teams to streamline workflow and improve data accessibility.

Helped develop a user management platform for an analytics company, providing secure account creation, role-based access control, and team management features. Integrated the system with Power BI to deliver interactive dashboards, enabling users to directly access and explore analyst-created reports within the platform.

Developed a next-generation chatbot to replace a legacy system, leveraging large language models for improved responses. Responsibilities included dataset preparation, fine-tuning and evaluating LLMs, and prompt engineering. Built the backend with FastAPI, containerized with Docker, and integrated MongoDB for conversation history. Implemented a vector database to enable Retrieval-Augmented Generation and applied zero-shot classification transformers for automated output evaluation.

Helped develop a chatbot designed to reduce the workload of human agents by automating routine tasks. The project involved analyzing KPIs with Tableau and SQL, performing text analysis with Python, and fine-tuning Transformer models to improve response quality and coverage.

Worked on a model to detect anomalies in time series data for use in a monitoring system. Analyzed and visualized datasets using Pandas, Statsmodels, and Seaborn, and built machine learning models for anomaly detection with Scikit-learn to identify unusual patterns and trends.

Developed and evaluated machine learning models to detect malware in Android applications, focusing on maximizing precision and recall. Built and compared multiple XGBoost models using different feature sets, analyzing performance with F1 score, precision, recall, confusion matrices, and precision-recall curves. The most comprehensive model, leveraging all available features, achieved strong results in identifying malicious apps.

Fine-tuned and compared DistilBERT and RoBERTa models for fake news detection using a labeled dataset of news articles. Evaluated model performance on multiple datasets with accuracy, precision, recall, and F1-score metrics, with RoBERTa achieving slightly superior results. The work culminated in a research paper published at the MIPRO Convention in Croatia.

Helped develop an open-source system, Logs2Graphs, to create graph representations of system logs for anomaly detection, log prediction, and root cause analysis. The system supports multiple publicly available log sources and is designed to be easily extended to additional sources, enabling structured visualization and analysis of log data.