February 2024 – July 2024

6 month internship at Adnovum in software engineering in Zurich, Switzerland. Working with spring boot microservices using agile software development methodology.

January 2015 – May 2023

For over eight years, I had the privilege of tutoring more than 20 students in mathematics, physics, and sciences. What began as a simple way to help classmates grew into a meaningful part of my life. Through sharing my passion for these subjects and watching my students gain confidence, clarity and curiosity I found meaningful joy. My approach was always guided by patience, compassion and this genuine wish to see them succeed, not just at school but also in life.

This project explores bandit algorithms for recommendation systems, from basic strategies like UCB to more advanced linear bandits. After testing on simulated data, we applied the models to real-world Amazon Music reviews. While bandits performed well in controlled settings, linear regression outperformed them in real-world prediction tasks. Still, bandits showed promise in data-scarce environments thanks to their adaptability and online learning capabilities. The study highlights when and where bandits can be effective tools in recommender systems.

This project tackles ternary sentiment classification (positive, neutral, negative) on 100K+ labeled sentences. We compare classical ML models to transformer-based architectures like RoBERTa and DeBERTa, and explore preprocessing, ensembling (softmax averaging, majority voting), and data augmentation using LLM-generated paraphrases. Our best model—an ensemble of fine-tuned transformers—achieved an L score of 0.9034 and a weighted F1 score of 0.83.

This project leverages the PhysioNet 2012 Challenge dataset to predict ICU patient mortality based on multivariate time-series data from the first 48 hours of ICU stay. We explore classic machine learning models (Random Forest, Logistic Regression), deep learning approaches (LSTM, BiLSTM, Transformers), and representation learning techniques. The study also investigates the use of LLMs for few-shot predictions and evaluates the performance of different feature aggregation methods. Key findings highlight the effectiveness of Random Forest for tabular data and the challenges of applying Transformers and contrastive learning to ICU time-series data.

Connect Out is a social activity platform that enables users to join and discover new activities based on their interests. It fosters social interaction and community engagement, making it easier to meet like-minded people and discover new hobbies.

A mobile application built using react native with expo cli and typescript, coupled with a backend in node and mongoDb. The mission of this project is to allow students to sublet their rooms during holidays.

Coded a social network application (with a friend) for our own and neighboring universities. Allowing students to express themselves about courses and campus life. Frontend coded in VueJS and backend in Java Spring.