As part of the Data Mining course at ITU Copenhagen, our group carried out a collaborative research project applying machine learning techniques to analyze real-world datasets. Together, we worked through the full data mining pipeline, from data collection and preprocessing to exploratory data analysis, modeling and evaluation. Our approach combined both supervised and unsupervised methods, including k-means clustering, decision trees and other classification algorithms implemented in Python.

Throughout the project, we investigated how different preprocessing strategies, such as feature engineering, normalization and dimensionality reduction affected model performance and interpretability. We benchmarked multiple models, tuned hyperparameters and validated results using cross-validation, confusion matrices and cluster quality metrics.

Our findings were consolidated into a comprehensive report that discussed our methodological choices, compared algorithmic strengths and limitations and reflected on the practical relevance of our results. The project strengthened our collective ability to design, implement and communicate robust data-driven analyses.

As part of the Othello GameAI project for the Introduction to Artificial Intelligence course at ITU Copenhagen, I co-developed a competitive Othello-playing AI using the minimax algorithm enhanced with alpha-beta pruning. The implementation involved recursive decision-making logic for both MAX and MIN players, with utility-based move selection and depth-first search. To optimize performance, we introduced cut-off mechanisms including a fixed search depth of 7 and a time constraint of 20 seconds per move, ensuring responsiveness while maintaining strategic depth. I contributed to designing and refining the core decision functions (MaxValue and MinValue), integrating utility evaluation and pruning logic to reduce computational overhead.

To guide strategic play beyond terminal states, we engineered a custom evaluation function based on five heuristics: legal moves, token difference, edge control, corner control and corner adjacency penalties. These components were weighted through empirical testing to maximize win rates against baseline AI opponents. The final utility function balanced positional advantage with long-term control, enabling our AI to consistently outperform simpler strategies and other group's AIs. This project sharpened my skills in algorithmic design, heuristic modeling and performance tuning within a collaborative development environment.

Designed, built and optimized a line-following robot equipped with a gripper arm, integrating both mechanical fabrication and embedded systems development. The project combined multiple prototyping techniques, including 3D printing and laser cutting for the chassis and structural components, as well as soldering and PCB design for custom electronics. I developed the full electronic system, from circuit design to assembly, ensuring reliable connections between sensors, actuators and the microcontroller. The robot’s mobility was driven by DC motors controlled through an H-bridge, while infrared sensors were calibrated and positioned for accurate line detection and navigation.

On the software side, I programmed the microcontroller (Arduino-based) to process sensor input in real time and adjust motor speeds for smooth line tracking. I implemented and tuned control strategies to balance responsiveness with stability, ensuring the robot could handle curves and intersections efficiently. The gripper arm was designed as a lightweight mechanism actuated by a servo motor, allowing the robot to interact with objects along its path. This required careful mechanical design and integration of actuators with the control logic. The final system demonstrated a complete mechatronic workflow covering mechanism design, electronics integration and algorithmic optimization and highlighted my ability to bridge hardware and software in a functional robotic prototype.

Image segmentation is a fundamental task in computer vision with applications in warehouse automation, inventory management and package sorting. This work investigates the application of pre-trained semantic segmentation models to detect and segment boxes, barcodes and plastic bags in images. We construct a unified dataset by combining four heterogeneous sources totaling 8,854 images with varying class annotations. We evaluate a MobileNetV3-DeepLabV3 model on this combined dataset, measuring performance using Intersection over Union (IoU) and Dice coefficient metrics. We fine-tune the model using class-weighted cross-entropy loss with early stopping based on validation performance. Our results demonstrate that the pre-trained model achieves excellent segmentation performance with a mean IoU of 0.9433 across all classes, with particularly strong results on boxes (0.9480) and the background class (0.9578). We analyze the challenges of learning from heterogeneous datasets with significant class imbalance and identify key strengths and limitations of the approach for practical deployment.

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.