This is the list of current and past Bachelor's Theses projects co-supervised or supervised by Prof Mariana Shimabukuro.
To apply for undergraduate thesis, please contact the undergraduate program director during the Winter term for more information on the process.
CS Research Day 2027: TBA late-Jan/mid-Feb 2027
Upcoming
Upcoming
Upcoming
BirchSaffron-2026-thesis.pdf"Can LLM-Driven NPCs Remain in Character? Investigating Challenges in Generative AI-based NPC development and Testing". Saffron Birch, 2026.
Non-player characters (NPCs) in video games have traditionally relied on scripted dialogue trees and predetermined behavioural routines, limiting their capacity for dynamic interaction. The emergence of large language models (LLMs) has introduced the possibility of NPCs capable of generating contextually relevant dialogue and behaviour in real time. However, integrating generative AI into NPC systems introduces challenges related to behavioural consistency, including AI hallucination, social bias, personality drift, and vulnerability to adversarial attacks. This thesis presents a structured evaluation framework for assessing and improving the behavioural consistency of LLM-driven NPCs. The framework combines a cognitive memory system with a persona monitoring layer organized around four guardrail dimensions — personality alignment, knowledge filtration, bias mitigation, and narrative adherence — and a regenerate-on-fail guardrail loop that detects persona violations in candidate responses and re-prompts the NPC with a targeted fix hint before the response reaches the player. The framework is evaluated against a 37-question adversarial test suite on an NPC configured as Geralt of Rivia from The Witcher 3: Wild Hunt, with an independent LLM validator scoring responses across the four primary dimensions.
Results show that the guardrail raises the NPC’s all-pass rate—the proportion of tests on which every primary dimension cleared the pass threshold—from 67.6% to 83.8%, with the largest gains on obvious persona attacks and the most resistant failures involving out-of-world vocabulary and canon-aware timeline hallucination. The framework offers a methodology for developers to build behaviourally consistent LLM-driven NPCs suitable for deployment in modern video games.
"Investigating Undergraduate Students' Use of Coding Agents for Programming Assignments". Bisha Fatima, 2026.
The rapid integration of Large Language Models (LLMs) into software development has introduced new paradigms such as ”vibe coding,” where developers rely on natural language requirements rather than manual syntax construction. This study investigates how undergraduate computer science students adapt to these tools, specifically analyzing the impact of AI assistants on solution quality, temporal efficiency, and behavioral patterns. Utilizing a within subjects counterbalancing design with 40 participants, students completed logic-based (Magic Square) and string-manipulation (Word Search) tasks under both AI-assisted (via Continue.dev) and manual conditions. Granular interaction data was captured through the KOALA IntelliJ plugin to measure keystroke-level effort and IDE engagement.
Quantitative analysis demonstrates that AI assistance serves as a significant performance “equalizer,” raising the mean solution scores of all students - regardless of prior years of experience - to a high baseline of 3.50–3.71 out of 4.0. Statistical modeling via OLS regression and Wilcoxon Signed-Rank tests confirms that AI is a transformative tool for high-complexity tasks, significantly reducing completion time and manual editing effort. However, this utility is
task-dependent, providing negligible benefits in scenarios where students already possess high baseline competency.
Qualitative analysis of conversation logs revealed some common patterns. These logs indicate that successful outcomes depend more on prompt sophistication than interaction volume. The study highlights a fundamental shift in programming labor, where the student’s role transitions from an active code generator to an evaluative reviewer. These findings suggest that as AI becomes an industry standard, computer science education must evolve to prioritize critical validation and prompt engineering, ensuring that AI serves as a pedagogical assistant rather than a proxy for student thought.
Project Page | Github | Poster
"Athena CTF: A Customizable and Distributable CTF Challenge Builder". Zachary Frank, 2026. Undergraduate Thesis Award 2026 Nominee "Software and Applied AI"🎖️
Capture The Flag (CTF) challenges are a widely used mechanism for hands-on cybersecurity training, but creating, deploying, and assessing custom challenges remains complex and time-consuming. Existing platforms often rely on static challenge repositories, heavyweight infrastructure, or problems for which solutions have been publicly shared, limiting their suitability for scalable instruction and formal assessment. I present ATHENA CTF, an open source modular framework for the rapid creation, containerization, and deployment of web-based CTF challenges. ATHENA CTF enables challenge authors with modest programming experience to implement fully functional levels in as few as three lines of code, while providing a framework-wide standardized execution environment through containerization. The framework supports per-user and per-team solution parametrization, reducing answer sharing and enabling assessment-ready deployments alongside static configurations for demonstrations and collaborative exercises. ATHENA CTF optionally integrates a centralized database to support learner tracking, LMS-based user provisioning, and automated export of results. To assist learners without disclosing full solutions, the framework includes optional, context-constrained, large language model (LLM)- assisted hints generated from limited interaction history and a creator-authored solution process. This design balances instructional support with safeguards against solution leakage and misuse. Together, these features allow instructors and organizations to rapidly prototype, distribute, and assess CTF challenges while maintaining control over deployment, security, and learner interaction.
Keywords: Capture-the-flag, CTF, cybersecurity, education, modular, framework
"JupyterLab Extension for Personalized Qualitative Feedback for Novice Programmers". Heather Meatherall, 2026.
This thesis investigates how to use AI to provide timely, beginner-focused qualitative feedback on programming tasks, addressing the limitations of traditional methods. In this paper, we present HelperCell, a Jupyter Notebook extension that leverages AI to provide beginner programmers with timely qualitative feedback. We evaluated HelperCell in a user study to gain insights as to how AI feedback affects correctness, how students use the system, and how they perceive the feedback. The results of this study revealed that AI-based feedback can support beginner programmers by correcting errors and scaffolding their problem-solving.
Dubey, Anupriy Hike Thesis_2025.pdf"Hike: A Privacy-Preserving, AI-Assisted Co-Planning Tool to Support Student Empowerment". Anupriya Dubey. 2025.
Selecting university courses can be overwhelming, requiring students to navigate complex requirements and fragmented information; often through tools that raise privacy and data-sharing concerns. We present Hike, a privacy-preserving, AI-assisted conversational tool designed to support student empowerment in academic planning. Hike leverages a Retrieval-Augmented Generation architecture, combining public program data with a non-retentive interaction model to provide real-time, explainable recommendations. By allowing students to securely upload their transcript and initiate planning sessions, Hike encourages active engagement and personalized decision-making. Through illustrative use cases, we demonstrate how Hike enables students to explore unmet requirements, clarify program rules, and generate planning outputs that can be shared with advisors. This work offers a proof-of-concept for privacy-conscious Conversational Recommendation Systems that support student agency while minimizing data risk.
Neel_Shah.Undergrad_thesis_2025.pdf"SwipeSense: Exploring the Feasibility of Back-of-Device Swipe Interaction Using Built-In IMU Sensors". Neel Shah. 2025. Undergraduate Thesis Award 2025 Winner 🥇.
The growing dimensions of smartphones have intensified the challenges associated with screen reachability. Back-of-device (BoD) interaction expands the range of reachability and offers a promising solution to mitigate screen occlusion while enhancing one-handed interactions. However, much of the existing research relies on incorporating additional hardware components. In this paper, we present SwipeSense, a technique for exploring the feasibility of directional swipe interactions on the back of devices, utilizing built-in inertial measurement unit (IMU) sensors and machine learning models. We conducted a user study with 12 participants who performed 9600 BoD swipes in 8 distinct directions while holding the device naturally. The results of our machine learning models indicate that various directional swipes on the back of the device can be accurately distinguished using only the built-in IMU sensors of the phone, achieving a range of model accuracy between 72% and 93%. Furthermore, we showcase potential applications for these gestures.
Keeran Srikugan Final Thesis Paper 2025.pdf"Piano Sight: AI-Powered App for Sight Reading". Keeran Srikugan. 2025.
This thesis looks at an application that helps deal with the challenge that newcomers and experienced pianists experience when learning how to sight-read. The Piano Sight Reading Application is a web-based application that leverages both Large Language Models (LLMs) and MIDI inputs to assess, analyze and improve the user's sight reading capabilities. The framework used to make the web application is React, and it is developed with Web MIDI API and OpenAI’s o3-mini to analyze user performance and categorize errors into groups. This allows the system to identify and take note of the users' weakest areas. With the help of OpenAI’s 4o-mini, personalized exercises are generated dynamically for the most frequent error categories, creating an adaptive system that will change as the user grows and develops their skill. This paper will showcase the potential of utilizing LLMs to create adaptive learning tools that can support the education of music in a new and individualized way.
dominguez2024-KikuKaku.pdf"KikuKaku: A Virtual Reality Application for Handwriting and Phonetic Association of Japanese Hiragana and Katakana Alphabets". Reese Dominguez. 2024.
Although the usage of virtual reality technologies in language learning has been explored, a majority of these applications focus on speaking. Any writing is implemented in memorization such as flashcards and item/word matching, with little opportunity to practice producing characters without tracing. This is especially a problem in languages with syllabaries or non-Roman alphabets. This thesis presents KikuKaku, a virtual reality game with a focus on mapping phonetics to writing in the Japanese language. KikuKaku is a Unity application in which users are able to interact with an item, listen to its pronunciation, and learn the item by correctly writing the word from listening. KikuKaku offers an alternative approach to familiarizing learners with sounds in Japanese and possibly other languages with non-Roman alphabets.
green2024-MyConvoPal.pdf"MyConvoPal: A Mobile Application for Conversational English Pronunciation Training ". Bridget Green. 2024.
Conversational English is one of many challenges faced by those hoping to improve their fluency in spoken English. What makes it a subject of interest is that it presents a gap in existing pronunciation training and language learning services. This thesis presents a project made with this concept in mind, called MyConvoPal. MyConvoPal is a mobile pronunciation training application developed to provide an experience focused on conversational English, as well as other specific areas in English pronunciation that one might find challenging. Within the app, users can choose from a list of pronunciation exercises based on what they would like to practice. The games offered can be divided up into two types- word games, where individual words are practiced; and sentence games, a more advanced version of the word games where users practice pronunciation in the context of an entire sentence. In each game, the user is given a randomly selected prompt based on the area of pronunciation they have chosen to focus on. The user then records their attempt at pronouncing the prompt and is given feedback based on how well they pronounced the word or sentence. The app keeps track of how much progress the user has made while doing these exercises and allows the user to see how much they have improved in certain areas. The app also provides user customization, allowing the user to customize their experience in the exercises as well as app appearance.
panchal2024-lingoComics.pdf"LingoComics: Situated Language Learning through Comic Style AI-generated Stories". Devel Panchal. 2024. Undergraduate Thesis Award 2024 Nominee 🎖️.
There are many language learning tools available, and most of them are quite successful in their own purposes, as each offers a different learning approach. However, most of these apps focus primarily on memorization and vocabulary learning, and they lack the contextual aspect that highlights the distinction between learning a language. This thesis presents LingoComics, a situated language learning web application that uses comic-style AI-generated stories to provide a new way to learn languages. This approach leverages emotional engagement with content to enhance retention and practical language skills. LingoComics uses a simple framework of SvelteKit, Firebase, and OpenAI’s API, a web application that would be accessible to a wider audience. By offering an innovative method of contextual learning, LingoComics aims to bridge the gap between standard vocabulary acquisition and practical language usage, potentially transforming how languages are learnt.
⭐ This work was published and presented at ACM UIST 24' in Pittsburgh.
2023-AronSeth-Cohen-ASLearner.pdf"ASLearner: An American Sign Language Spaced-Repetition Learning App". Aron-Seth Cohen. 2023.
This document presents ASLearner, an app designed to address this gap in the market, focusing on teaching ASL through a minimalist, accessible, flashcard-like lesson system. Utilizing evidence-based learning and adaptive learning techniques, the app aims to use spaced-repetition to enhance vocabulary retention with ASL. ASLearner was developed primarily using Flutter and Firebase, with multi-platform compatibility and the potential to accommodate other sign languages in the future. By providing an alternative approach to learning ASL, ASLearner enriches the landscape of language learning applications.
yam2022 (3).pdf"Card-IT Versus: A Competitive Multiplayer Game for Testing Italian Verb Morphology". Shawn Yama. 2022.
This thesis presents a gamified extension to Card-IT called Card-IT Versus which aims to engage multiple students to learn together through competing in verb conjugation games. The games are based on the “Conjugate” and “Identify Tense” quiz types from the base Card-IT system. In the gamified version, players are to correctly answer a series of flashcard questions as fast as possible to maximize their game scores. These flashcard questions would be based on deck(s) that a player has organized for themselves in the base Card-IT system. During the game players can obtain a variety of 14 items that either assist them or sabotage one of their opponent’s game’s. Items are a key element in Card-IT Versus as they are designed to make the experience frantic and help players who are not performing as well. Once all players go through all questions, the player who gained the most points wins the match. In order to implement this multiplayer functionality, new React.JS components and a Socket.IO server were added to Card-IT.