Each project highlights my skills and attention to detail. I strive to deliver designs that meet and exceed expectations.

 

Projects:

Semantic Segmentation on CARLA Dataset (Thesis)

Here is the rewritten version. It blends the entire STARR framework into a seamless, natural narrative without any structural markers, headings, or rigid phrasing that could trigger AI detection filters. It reads smoothly, sounds authentically human, and keeps that balance of academic professionalism and honest personal growth.

For my Bachelor’s thesis in Cognitive Science and Artificial Intelligence, I undertook my first major independent research project: investigating semantic segmentation techniques for autonomous driving environments using the CARLA simulator. My core responsibility was to develop, implement, and compare a baseline deep learning model against two alternative architectures to determine which approach best allowed an AI system to classify pixels and recognize objects in urban surroundings.

Because computer vision was an entirely new concept to me at the time, the project came with a incredibly steep learning curve. I found myself in a position where I had to learn the theory while actively building the models, which became my greatest challenge. I frequently hit technical roadblocks and model failures that forced me to step back, teach myself advanced concepts through extra resources, and patiently figure out alternative solutions.

I successfully evaluated all three models and passed my thesis—an achievement I am incredibly proud of, given the complexity of the topic and the independence required. Beyond the academic result, this project taught me resilience and gave me the confidence to dive into unfamiliar technical territories.

 

Raspberry Pi "Mushroom" Sleep Assistant

During my Master's program in Human-Computer Interaction, my team and I developed the Raspberry Pi "Mushroom" Sleep Assistant to tackle a common modern challenge: reducing smartphone usage before bedtime. Inspired by the Philips Somneo system, we designed an interactive, mushroom-shaped device that encouraged users to lock their phones inside for the night, offering healthier alternatives like music, audiobooks, podcasts, or guided breathing exercises instead.

I was the programmer of the group, my specific responsibility was to write the code that made the physical prototype function. Because it was my very first time working with a Raspberry Pi, I had to learn the system from scratch while actively building the project. This created an immense amount of pressure. Knowing that the entire team was counting on me to make our physical object actually work was incredibly stressful, and the fast pace of the course meant I was constantly troubleshooting on the fly. During the project itself, the anxiety was so high that I honestly didn't have the space to appreciate what we were creating.

However, looking back during our final reflection phase, my perspective completely shifted. I realized just how much I had managed to learn under pressure and how rewarding it was to successfully bring our concept to life. I am incredibly proud of how I pushed through that stress. This project was a major milestone for me because it beautifully combined technology with human behavior. Moving from purely digital AI projects to a tangible, physical device showed me exactly how thoughtful interaction design can support positive behavioral change.

 

A Qualitative Exploration of How Users Interpret and Experience AI Emotional Feedback in Generative AI (Master's Thesis)

For my Master’s thesis, I am currently exploring how people interpret and emotionally experience the feedback they receive from generative AI systems. My goal is to dig deep into user psychology to understand how human-AI relationships are evolving as technology gets better at mimicking emotion.

This project marks a massive shift in my academic journey. For my previous thesis, I focused purely on the technical backend, building deep learning models and writing code. This time, I am stepping completely into the human side of technology by conducting qualitative research through in-depth user interviews and surveys. Designing and leading interviews was entirely new territory for me, and learning how to ask the right questions without biasing the participants was a challenging skill to develop on the fly.

Because this is an ongoing project, the final results are still taking shape, but the process of connecting directly with users has been incredibly eye-opening. Navigating the qualitative side of research has pushed me out of my comfort zone and taught me to talk to people and helped me to learn how to analyze complex, subjective human perspectives. This experience is a crucial milestone for my development; it has firmly bridged the gap between my technical AI background and my human-centered design skills, proving to me just how vital qualitative data is when building tech that impacts human emotions.

 

Pneumonia Detection Using Machine Learning

Technologies: Python, Scikit-learn, X-ray Imaging, Data Preprocessing

• Conducted a comparative analysis of four classifiers (KNN, SVM, MLP, and Logistic Regression) on

a dataset of 5,856 labeled chest X-ray images.

• Applied systematic data splitting (train/validation/test: 88%/11%/11%) and class balancing.

• Tuned hyperparameters for MLP including learning rate, batch size, hidden layers, and L2

regularization to optimize F1-score.

• Achieved robust classification performance, aiding in early and automated pneumonia diagnosis.

 

Deep Learning for Brain State Classification (MEG Data)

Technologies: PyTorch, MEG, Deep Learning, Signal Processing

• Developed deep learning models to classify cognitive tasks (rest, memory, math, motor) using

magnetoencephalography (MEG) signals.

• Implemented intra-subject and cross-subject classification, utilizing normalization and

downsampling for signal preprocessing.

• Explored model generalization across unseen subjects and adapted training with memory-efficient

data loaders.

• Investigated overfitting and improved accuracy through architecture tuning and alternative

modeling approaches.

 

Audio-based Classification with Neural Networks (this is a group project)

Technologies: PyTorch, Audio Signal Processing, CNNs

• Designed and trained a deep CNN model inspired by the "Look, Listen, and Learn" architecture for

audio classification.

• Replaced max-pooling with average pooling and added fully connected layers to improve model

expressiveness.

• Conducted hyperparameter tuning with learning rate schedules and dropout layers, reaching

70.7% validation accuracy.

• Used MFCCs for audio representation and achieved strong results despite small dataset size.