Eight MSc theses to be presented on Wednesday June 4, 2025

Wednesday June 4 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Eight theses will be presented.

You will find information about how to follow along under each presentation. There will be presentations in two different rooms: E:2116 and E:4130 (Lucas). See room for each presentation. A preliminary schedule follows.

Please note that there will also be thesis presentations on Wednesday June 11, schedule at: (Link to follow shortly)

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.

10:15-11:00 in E:4130 (Lucas)

Presenters: Ludvig Eskilsson, David Pettersson
Title: PriceFinderAgent: A Purpose Built AI Agent to Extract Phone Contract Prices
Examiner: Mathias Haage
Supervisors: Pierre Nugues (LTH), Richard Toth (Prisjakt AB)

Web scrapers are programs used to collect structured data from the public web, but traditional scrapers often break due to hard-coded logic and changes in site structure. Recent advances in generative AI through large language models (LLMs) and large multi-modal models (LMMs) offer new opportunities for building more robust alternatives. This thesis explores the use of AI to automate web scraping workflows. We focus on extracting phone subscription contracts from three Swedish mobile carrier websites, a domain that requires both reasoning and interaction due to the complexity of contracts and layouts. We evaluate multiple AI-assisted strategies, including script-based automation and agentic frameworks. Among these, a multi-agent LMM-based system using vision achieved the strongest results, reaching 90% precision, 100% recall, and 92% accuracy. The findings show that modern LMM agents can serve as reliable scrapers for dynamic websites, provided they are guided with effective prompts and supported by appropriate navigation tools. 

Link to popular science summary: Link to be added

11:15-12:00 in E:4130 (Lucas)

Presenters: Tim Julius Wikdahl Thunström, Johannes Gerding
Title: Graph Neural Network-based approach to leak detection in the water industry
Examiner: Pierre Nugues
Supervisors: Leonard Papenmeier (LTH), Jakob Folkesson (Backtick Technologies AB)

Efficiently detecting leaks in water distribution networks is crucial for assuring access to clean water. Modern sensor equipment allows for increased automation of this process but also poses the need for good algorithms to leverage the sensor data successfully. In this study, we analyze the effectiveness of using graph neural networks (GNNs), a type of neural network designed to handle graph-based data, for leak detection on a synthetic dataset. To do this, we utilize GNNs from a graph outlier detection framework called PyGOD as well as custom GNNs implemented in PyTorch Geometric. These models are then compared to a popular heuristic method called minimum night flow. We find that the models in PyGOD fail to improve upon the heuristic while the custom GNNs achieve improved detection performance on some metrics. Furthermore, we highlight the importance of standardizing the evaluation of models in this domain.

Link to popular science summary: Link to be added

13:15-14:00 in E:4130 (Lucas)

Presenters: Linnéa Gedda, Ella Hammer
Title: Identifying Surgical knot tying with Machine Learning
Examiner: Volker Krueger
Supervisor: Maj Stenmark (LTH)

The purpose of this study was to research different machine learning models and their ability to identify knot-tying specifically during open-heart surgery. Four different models were examined: the SV-RCNet (a ResNet followed by an LSTM), a 3D-CNN, a Transformer and a Decision Tree, using different hand tracking methods. MediaPipe hand landmarks, and a Yolov8 trained specifically on surgical data, which produced segmentation masks and bounding boxes. The models were trained to classify data between "knot tying" and "not knot tying", where the non-knot tying data was an assortment of different surgical actions performed during surgery. The data used was videos recorded during real-life surgeries, which was run through the different hand tracking models.

Link to popular science summary: Link to be added

N.B. This thesis will be presented on June 11 instead

Presenters: Joel Pistora, Gustav Kristiansson
Title: Enhancing the Prioritisation of Security Requirements in Software Development: A Cost-Effectiveness Perspective
Examiner: Björn Regnell
Supervisors: Alma Orucevic-Alagic (LTH), Gustav Lundsgård (IKEA)

A major issue in cybersecurity is the lack of relevant knowledge for prioritising security requirements due to the diverse context of software systems. A cross-team priority order if often ineffective, highlighting the need for a method that accounts for system-specific context parameters when prioritising security requirements.

This study was conducted at a large-scale organisation, with the goal to investigate how to formulate a contextual requirement prioritisation method. Through a workshop involving cybersecurity experts, key system parameters that influence prioritisation decisions were identified. These parameters were used as context for a survey in which respondents made pairwise comparisons of security requirements. The collected data was analysed using a developed scoring model.

The results indicate that prioritisation changes based on system context and that respondents reached general consensus. Based on these findings, a six-step method was proposed to support the contextual prioritisation of security requirements in practice.

Link to popular science summary: Link to be added

14:15-15:00 in E:4130 (Lucas)

Presenters: Georg Lundqvist, Daniel Baldrigue Andrade
Title: Intelligent Robotic Depallitizing Systems for Blank Feeding
Examiner: Jacek Malec
Supervisors: Volker Kreuger (LTH), Daniel Cederström (Tetra Pak)

Manual depalletizing of paper-wrapped Tetra Pak boxes remains a labor-intensive process in otherwise automated lines. This thesis develops and validates an intelligent robotic depalletizing system, advancing previous work from TRL-4 to TRL-5 by demonstrating operation in a factory setting. A collaborative robot UR10e with a Robotiq PowerPick is driven via ROS2; perception fuses Grounding DINO, Segment Anything, and an Intel RealSense D435i depth camera to localize boxes and predict grasp planes in real time. Task execution is organized as a SkiROS2 behavior tree, which allows adaptive skill execution. Factory trials depalletized 80 boxes at an average of 70 seconds each at a rate of 51 boxes per hour, exceeding the requirement of 30 boxes per hour, and achieved the completion of the 90 % cycle without human intervention. The results confirm the viability of vision-language-guided robotic depalletization and outline the steps toward full deployment while highlighting the remaining challenges around small deformable boxes and localization. 

Link to popular science summary: Link to be added

14:15-15:00 in E:2116

Presenters: Elsa Cervetti Ogestad, Julia Karlsson
Title: Compiling automation programming languages to WebAssembly
Examiner: Ulf Asklund
Supervisors: Görel Hedin (LTH), Jonas Bülow (Schneider Electric)

WebAssembly is getting more and more popular as a portable platform for embedded systems.. This thesis introduces a general approach to developing a compiler for languages similar to the IEC-61131-3 defined Structured Text and Function Block Diagram, targeting WebAssembly. WebAssembly enables program compatibility between standard PC hardware and resource-constrained embedded systems, and allows a shared runtime environment across the layers of building automation systems. 

We present two alternatives: a combination of generating LLVM IR, optimising it and using Clang, or generating C++ and using Emscripten. These are evaluated through a case study of two of Schneider Electric's languages, where it was shown that utilising LLVM directly, rather than through Emscripten, provides shorter and more consistent compilation times. Due to the associated heavy development effort of LLVM, Emscripten could be a good alternative if compilation time is not a priority, as they both generate equally fast executables in WebAssembly.

Link to popular science summary: Link to be added

15:15-16:00 in E:2116

Presenters: Oliver Persson, Erik Neman
Title: Refactor Me If You Can: When AI Rewrites The Mess 
Examiner: Görel Hedin
Supervisors: Andreas Bexell (LTH), Emma Söderberg (LTH), Roger Dreyer (ASSA ABLOY)

Refactoring is a crucial aspect of software engineering. As Large Language Models (LLMs) have shown promising capabilities in all software-related tasks, developers can now utilise LLMs for automatic refactoring. 

Our work investigates the code refactoring capabilities of LLMs in improving the maintainability of a large codebase, by evaluating several models. We experiment with different input partitioning strategies, such as class-level and method-level refactoring, and evaluate how different reasoning levels affect the refactoring performance. 

Our results show that LLMs generally improve the maintainability of code. 62.1\% of the attempted refactorings were successful, with smaller code snippets showing a higher success rate than larger ones. 

Method-level refactoring achieves better results than class-level refactoring. However, method-level refactoring is more expensive and time-consuming. 

We find no significant difference in how different reasoning levels affect the maintainability of code. Furthermore, a high level of reasoning is significantly more expensive and time-consuming than low reasoning levels. 

Link to popular science summary: Link to be added

16:15-17:00 in E:2116

Presenters: Johan Hummel, Ella Viirola
Title: From C 2 Rust: Evaluating the Feasibility of Translating C to a Memory-Safe Programming Language at a Large Organization
Examiner: Ulf Asklund
Supervisors: Emma Söderberg (LTH), Peter Kornevi (Ericsson AB), Henrik Eideberg (Ericsson AB)

The C programming language, while offering high performance and low-level control, is memory-unsafe. This lack of safety makes it prone to programming errors that can result in serious software vulnerabilities and system instability. This thesis investigates the feasibility of transitioning Ericsson's extensive C codebase to a memory-safe alternative, specifically Rust. The study explores various aspects of such a transition, including compatibility with Ericsson's existing compiler toolchain, to what extent automatic translation with the C2Rust tool is possible, the performance impact, as well as developer attitudes toward adopting Rust. In addition to the analysis of Ericsson’s internal codebase, two open source projects were also studied.. The purpose of this was both to deepen the understanding of translation challenges and to enable more controlled performance measurements, which proved difficult on the internal codebase. 

Link to popular science summary: To be added