BS CS Capstone Papers

Mathematical learning is an integral part of computer science. Through the years of undergraduate Computer Science, students are introduced to a large number of mathematical concepts, some of which are not intuitively visualizable, particularly when traditional teaching relies heavily on static representations. To address this, we developed an interactive web-based platform that visualizes fundamental mathematical ideas dynamically and accessible. Employing our combined expertise in mathematics and software development, the platform integrates modern frameworks, including ASP.NET Core for server-side architecture and React for front-end interactivity. We utilized the Manim library in Python to dynamically generate mathematical animations. Using the tool, the users will be able to not only learn, but also save and share mathematical visualizations with other users. The goal of the project is to deploy a learning environment for students that will bridge the gap between abstract mathematics and intuitive understanding.

The diary blog platform is a secure, AI-enhanced web application that enables users to write, manage, and analyze personal journal entries. Designed with user privacy and personal reflection in mind, the system provides a seamless registration and login process secured by verification codes to ensure account authenticity. Once authenticated, users can create, edit, and organize diary-style posts using a robust text editor that supports rich formatting options such as headings, hyperlinks, and other styling tools, allowing for expressive and structured entries. In addition, users can like and comment on posts, and receive notifications about new interactions. Beyond traditional blogging capabilities, the platform integrates artificial intelligence to provide meaningful insights into the user’s writing. An AI engine analyzes each entry to detect emotional tone and mood, and suggests improvements to enhance clarity and self-expression. This allows users to reflect more deeply and observe patterns in their emotional state over time. A personalized dashboard complements this experience by visualizing key metrics such as post frequency and mood trends, offering a broader perspective on journaling habits. While all user-generated content is stored securely and accessible only to authenticated users, posts can be marked as either private or public. Even for public posts, full access is restricted to registered users, keeping a controlled sharing environment. A publicly accessible home page organizes post summaries by category, allowing visitors to explore post titles and previews without viewing full content. In summary, the platform integrates secure account management, personalized creative expression, and intelligent feedback to deliver a digital journaling experience that is both personal and insightful. It demonstrates the power of combining modern web technologies with AI-driven analysis in an innovative and thoughtful way.

Linear Programming is a viable tool for solving optimization problems. With almost eight decades of history and an extensive theoretical and practical background developed by mathematicians and computer scientists over this time frame, it is applicable to a wide range of real-world problems. There are many linear and integer linear programming solvers which were developed with the purpose of solving any linear programming problem as efficiently as possible. With so many solvers available from both open-source and commercial communities, one might raise a question regarding which one of these LP solvers is the most feasible option in terms of efficiency, ease of use, available resources, and other criteria. One might also be interested in how a given real-world problem can be efficiently converted to a linear program. This report aims to examine several linear programming solvers from different origins and compare them using various real-world optimization problems, while also showing how each problem is modeled as a linear program and what results are obtained in each case. The outcomes should give the reader insights about linear program modeling and which of the on-hand tools is the best option in a given setting.

This project presents a low-cost, webcam-based eye-tracking communication system designed for individuals with severe motor impairments, such as those affected by locked-in syndrome. The system enables users to select among five directional gaze inputs—left, right, center, up, and down—confirmed through an intentional blink. Horizontal gaze is determined using a dark-pixel ratio between the left and right halves of the eye, while vertical gaze is detected using Eye Aspect Ratio (EAR) thresholds. Blink confirmation is implemented using blink duration filtering, with a pre-blink label retrieval mechanism to avoid misclassification caused by pupil movement during blinking. Thresholds for each gaze zone were empirically adjusted based on testing under various lighting and distance conditions. Data collected from manually labelled sessions were used to measure system accuracy. Preliminary results demonstrate high reliability for most gaze directions, with ongoing refinements aimed at improving downward gaze classification. Future extensions include predictive text integration, text-to-speech output, and adaptive machine learning-based gaze calibration to enhance usability and responsiveness.

The performance improvement of large language models (LLMs) mainly came from scaling pretraining. However, a new scaling paradigm emerged called test-time compute, which uses more computation at the inference time of language models to get better results. There have been extensive works suggesting various test-time compute scaling strategies, but the derivation of laws (in functional form) describing the scaling dynamics of these methods is still an underexplored research question. In this capstone project, we try to bridge this gap. We consider evolutionary algorithms as a means for scaling language model test-time compute. We propose a parametric law for modeling the expected performance of language model-powered evolutionary algorithms depending on the language model parameters (N) and the number of evolutionary algorithm iterations (k). We show that in molecular optimization tasks, our law is able to accurately extrapolate in N and k. Moreover, we find a tradeoff between language model size and the number of evolutionary algorithm iterations.

This study investigates the generation of realistic human faces from textual descriptions for use in law enforcement, particularly suspect identification. We curate a demographically balanced subset consisting of 10,000 images from the Illinois DOC Labeled Faces dataset. We also introduce a fully automated annotation pipeline that uses a multimodal large language model to generate descriptive captions. We hypothesize that data quality, including alignment, and demographic balance, significantly improve the performance of pretrained generative models. Experimental results using state-of-the-art text-to-image models support this. These findings highlight the importance of task-specific datasets for accurate and reliable AI-generated facial imagery in investigative contexts.

Large Language Models have transformed how we interact with information, yet they can hallucinate and struggle with complex, multimodal inputs. Retrieval-Augmented Generation (RAG) improves factuality by grounding models in external data, but traditional RAG still falls short on visually rich documents and complex, multi-step questions. To overcome these limitations, we present Activeloop-L0, a reasoning-augmented retrieval pipeline designed for open-domain Question-Answering (QA) over a large set of user-uploaded multimodal documents. At query time, Activeloop-L0 dynamically interleaves retrieval and inference: it decomposes complex questions into subqueries, refines each search result using a vision-language perception module, and applies lateinteraction multimodal retrieval to surface the most relevant passages from text, tables, charts, and diagrams. By iteratively filtering and compressing results based on the user’s intent, our system achieves 85.6% QA accuracy on a visually rich data benchmark, setting a new state-of-the-art.

Super-resolution microscopy techniques have enabled imaging at the nanometer scale using fluorescent molecules (fluorophores), overcoming the diffraction limit of conventional light microscopy. Among these techniques, single-molecule localization microscopy (SMLM) achieves high spatial resolution by sequentially capturing numerous frames to localize individual fluorophores. However, this process results in low temporal resolution, often taking several minutes to generate a single image, which hinders the ability to study dynamic biological processes. This work aims to address this limitation by exploring deep-learning-based spatiotemporal interpolation methods to reconstruct videos from raw SMLM frame sequences, which are typically summed into static images. We explore hybrid architectures combining convolutional neural networks for spatial feature extraction and recurrent neural networks for temporal feature extraction, reconstructing the videos and enabling the visualization of nanoscale dynamics.

Artificial Intelligence today has the power to generate realistic text, images, and videos. However, editing existing videos—particularly changing specific objects while keeping the visual consistency of the video—remains a nontrivial challenge. This research investigates the use and adaptation of generative models to build a system for automated object replacement in videos. The core challenge lies not only in visually replacement but also in ensuring that the introduced object inherits the original object's motion patterns and interactions within the scene. Rather than building novel models from scratch, this work focuses on reusing and augmenting pre-trained architectures, both to promote methodological efficiency and to alleviate the carbon footprint issue associated with extensive GPU usage. This research aims to develop an ecologically mindful, motion-aware object replacement system for videos, with promising applications in intelligent video editing, film post-production, augmented reality, experimental data generation, and event reconstruction.

The past is never silent. It speaks to us in many ways, often quite literally, through handwritten manuscripts or carved stone inscriptions. Armenian monuments are particularly rich in the latter. These inscriptions serve as vital records of cultural and linguistic heritage, offering insights into the lives, beliefs, and traditions of Armenians during the Middle ages. However, reading, comprehending, and even detecting these inscriptions pose significant challenges. Due to weathering, vandalism, erosion, and the complexity of ancient scripts, many of these texts remain unreadable. Yet, the few existing studies indicate that deciphering these messages from the past is feasible with technological advancements. The modern inscription recognition methods are dominated by a two-phase approach – a character detection followed by their classification. Most notably, an end-to-end pipeline for reading Armenian stone inscriptions has been recently developed based on visual transformers, with the scarcity of available training data highlighted as the major limitation. In the present project we study a unique, newly created and unexplored collection of digital twins of Armenian tapanakars (tombstones) and khachkars (cross-stones). We process both 2D images and full 3D models aiming at the improvement of the detection accuracy. In the current stage we focus on hierarchical segmentation of the images using the detected geometrical and statistical features. The segmentation is implemented at three levels – multi-line texts, individual text lines and characters or strokes in the text line. Next, we apply the results to character classification and estimate the accuracy of the generated character images. Since the detection stage of the algorithm is universal for any kind of shapes, it opens up new research avenues that extend beyond text recognition alone. The same pipeline can be adapted to identify decorative motifs, geometric symbols, and other visual patterns commonly found on tapanakar surfaces.

Animating 3D characters is often labor-intensive, requiring careful rigging and masking of models to prepare them for animation. Many game developers lack the specialized skills needed for rigging, while animators aim to minimize the time spent on this repetitive task. This project aims to automate the process of making arbitrary humanoid 3D models ani- matable. We propose a data-driven method that uses graph neural networks to automatically rig and mask characters in the widely-used Mixamo format. Our method explores and compares different design choices. Specifically, it examines incorporating priors about joint locations by predicting them in 2D space versus relying purely on geometric and topological cues. We find that including joint-based priors can significantly improve the accuracy and robustness of the rigging process, particularly for models with unusual proportions or noisy geometry.

Spectral imaging has been a reliable tool for terrain analysis, giving valuable information about soil quality, vegetation type, and much more. Only recently has its potential been discovered in a broader range of fields such as biomedical imaging, the manufacturing sector, art restoration, and forensics. With the introduction of other fields, automating the processing of huge information obtained from spectral imaging has become necessary, making machine learning a common tool used for working with spectral data. However, gathering enough data for a model to successfully train on raises the need for new methods of spectral image acquisition or generation, especially in data-scarce or cost-sensitive contexts. Currently, spectral image generation is a poorly explored domain, and most of the works related to it are tailored to work with terrain data. In this report, a novel method for spectral image generation is proposed which is targeted for a hyperspectral fruit image dataset. More specifically, this report combines and adapts conventional RGB image processing, augmentation, and generation techniques for a new experimental method for working with and generating spectral images. Using the nature of hyperspectral images, the proposed methods boosts and optimizes the training process for image generation.

We propose an automated data-driven seasonal color profile classifier that generates personalized palette suggestions by extracting color-based features from face images. [1]. The project aims at the implementation of a virtual makeup try-on tool that will allow users to apply the recommendations in real time. This research offers a scalable solution for personalized color analysis, bringing image processing, computer vision, color theory, fashion, and cosmetic science together to support individuals in making more confident and color-informed choices in personal style and makeup. It is addressed in four directions: Detection of skin-related colored pixels in different color spaces using already-established color theory and image processing methods to enable the segmentation of different skin regions across varying lighting conditions; Formation of a skin-specific color palette using the extracted skin regions and identification of non-skin facial features such as the eyes, lips, and hair. These facial components are separately color-clustered to form a unique palette representing that individual’s natural coloring. Based on the observed color patterns, such as the skin undertone (cool, warm, neutral), the contrast level (high, medium, or low) of facial features, and the clarity (clear or soft), the person’s seasonal type (e.g., Soft Summer, True Autumn, Deep Winter, Clear Spring) is identified; - Finally, the seasonal types are mapped to makeup products to generate personalized recommendations and best-suited clothing color solutions.

Analyzing malicious programs deployed during cyberattacks is crucial for decoding attacker’s methodologies and developing proactive defense. Malware analysis, while invaluable, is a slow process - particularly in Linux environments where automated analysis capabilities lag significantly behind their Windows counterparts. Current solutions addressing the Linux ecosystem typically suffer from limited kernel visibility and inconsistency across Linux distributions. This paper presents Maltrace, a malware analysis system built on extended Berkeley Packet Filtering technology. It works by tracing the execution of malicious programs at the Linux kernel level, collecting a large amount of data about malware’s behavior, and then using analysis engines to turn this raw data into practical insights about the malware’s actions and goals.

The emergence of Large Language Models has been instrumental in the automation and advancement of numerous fields ranging from Natural language processing and Healthcare to Biological Structure Prediction tasks. However, the LLMs have been predominantly limited to English and have limited development in low-resource languages. This is primarily due to the lack of datasets for pre-training or fine-tuning. This work explores the possibility of generating and validating synthetic data for low-resource languages through the existing LLMs. We particularly focus on the development of Question Answering datasets as they have been shown to be the fundamental component for achieving human-like performance on numerous tasks and the principal evaluation benchmark for the state-of-the-art (SOTA) LLMs. We demonstrate the effectiveness of our approach by generating synthetic data for the Armenian language. The dataset allows us to evaluate and compare the zero-shot performance of the SOTA LLMs in the Armenian language. We perform a complete analysis of the generated syntactic data to show that the generated data is of high quality and can be used to bootstrap the development of NLP models for low-resource languages.