The goal of this project was developing a chatbot which utilized a Large Language Model which would replace an older chatbot system. The work on this project included analyzing and preparing datasets for fine-tuing LLMs, fine-tuinging and evaluatign LLMs, prompt engineering, backend development with FastAPI, continerization with Docker, setting up a vector database for Retreival Augmented Generation, setting up a MongoDB for keeping historical context, and using zero-shot classification transformers for evaluating the LLM output.

The goal of this project was developing a chatbot which would lessen the workload of human agents doing the same job as the bot. The project included analyzing KPIs with Tableau and SQL, analyzing textual data with Python, as well as fine-tuning Transformer Models.

The goal of this project was to develop a model for detecting anomalies in time series data, which would be used in a monitoring system. The project included analyzing and visualizing data with Pandas, Statsmodel and Seaborn, and building ML Models for anomaly detection using Scikit-learn.

The omnipresence of Android devices and the amount of sensitive information kept in them makes detecting malware in Android applications crucial. In this paper, the efficacy of using machine learning models for the purpose of malware detection in Android applications was examined, and several XGBoost models were developed and compared - each with a distinct feature set. We used the f1 score, precision, recall, confusion matrices, and precision-recall curves to compare the models. Accuracy was not considered since we needed a balanced dataset. One of the models we developed, which used all the available features in the dataset, had encouraging results with high precision and recall.

The goal of this project was to fine-tune two transformer models, namely DistilBERT and RoBERTa, and compare their effectiveness in fake news detection. Both models were trained on a labelled dataset of news articles and evaluated on two datasets, comparing their performance in terms of accuracy, precision, recall and F1-score. The results of the experiments showed that both models perform well, with RoBERTa achieving slightly better results overall. This project resulted in a paper that was published at the MIPRO Convention in Croatia.

The goal of this project was to develop a system which could create graph representations of system logs. These graph representations could then be used in logs anomaly detection, logs prediction, and root cause analysis guided by logs. Therefore in this paper, we present logs2graphs, an open-source system for the creation and visualization of such graph representations of log messages, which is compatible with several publicly available log sources and expandable to other log sources.