University Honors

Well-child visits (WCVs) are central to pediatric preventive care, yet immigrant families in the United States experience disparities in access and quality of WCV care. Structural barriers, language discordance, and cultural incongruence contribute not only to missed visits but also to increased caregiver burden in care management. In this scoping review, we synthesized the evidence on barriers to and interventions to improve WCV access among immigrant families. Using a scoping review approach, we identified studies focused on WCV access among immigrant and underserved families. Studies were included if they addressed structural, cultural, and language barriers that influenced caregiver burden and access to WCVs. Nine studies met the inclusion criteria, where we then synthesized qualitative, randomized controlled, and other quantitative studies. Across the studies, we found that language and cultural differences emerged as common barriers for immigrant families. Even when families were able to attend WCVs, the quality of the visit still depended on communication quality, cultural alignment, and trust in the information and care provided. Many preventive care models were unfamiliar to first-generation parents, which led to uncertainty and dissatisfaction even after completing visits. In underserved communities, children experiencing overlapping barriers such as poverty, limited English proficiency, and insurance instability were less likely to report regular physician visits or an established source of care. In this review, we identified many barriers to access to WCVs that could be addressed in health systems to improve care, including navigating unfamiliar health systems, translating medical information, coordinating insurance, and reconciling collective family decision-making norms within an individualistic care model. Advancing access to well-child visits in underserved regions requires culturally responsive, equity-focused strategies that address structural and communication barriers within pediatric preventive care systems.

Homelessness continues to be a major social issue in the United States, and one practical challenge is that people who need help often have difficulty finding clear and reliable information about available services. Food assistance, shelter programs, mental health support, and other community resources are often listed across different online directories, but these directories can be difficult to search quickly. This capstone project developed a retrieval-based artificial intelligence system for organizing and searching homeless-support and community resource information. The system was built in Python using structured service data stored in CSV format. Service records were represented as subjectñrelationshipñobject triples, allowing each providerís name, service category, address, phone number, website, ZIP code, and hours of operation to be stored in a consistent format. The system used sentence-transformer embeddings and FAISS similarity search to match user questions with relevant service records. It also used ZIP code and city-based location cues to prioritize services relevant to the userís requested area. Testing showed that the system could return multiple structured service recommendations for queries involving food assistance, location-based searches, and day-of-week availability. Instead of giving a single answer, the system presented several matching services with practical details such as address, phone number, website, and operating hours. Although the system depends on the accuracy and completeness of the underlying dataset, the project demonstrates how retrieval-based AI methods can make community resource information more organized, searchable, and accessible.

This project outlines the development of a journaling application designed to enhance deep self reflection with LLMs. This application addresses the existing privacy and ethical risks prevalent in existing cloud-based journaling chatbots. The app is built as a native macOS application using Flutter for the frontend graphical user interface and Python for the backend engine. OLLAMA to securely host local large language models (LLMs) and integrated into the backend through Langchain. Local storage is managed through SQLite for data serialization and FAISS vector stores for Retrieval-Augmented Generation (RAG). The application guarantees absolute data privacy with zero external network dependencies. The system functions as a reflective mirror following the principles of Acceptance and Commitment Therapy (ACT). Core feature capabilities include value systems analysis, temporal cognitive tracking, and contradiction awareness. Through natural language processing, the AI extracts virtues and behavioral metrics from daily entries. This data is fed through the RAG pipeline to generate highly personalized prompts that highlight past versus present value-behavior inconsistencies. The goal of prompting is to enhance user introspection without generating clinical advice through the LLM. The future goal of this application is to serve as a therapeutic aid for clinicians to analyze the progress of patients.

Accurate short-term equity forecasting is challenging due to nonstationarity and noisy market signals. This project develops an end-to-end forecasting prototype that predicts short-horizon price movement (future return over a user-selected horizon) for a selected stock. Historical openñhighñlowñcloseñvolume (OHLCV) data are retrieved using yahoo finance API and transformed into a compact set of core technical indicators, including daily returns, moving averages, RSI, Bollinger percent-B, average true range (ATR), and volume/range-based measures. To incorporate qualitative market information that may not be fully captured by prices alone, the system also retrieves recent news headlines for the chosen ticker and computes VADER sentiment scores, producing a sentiment signal used to generate sentiment-adjusted forecasts for demonstration. The forecasting model is a Long Short-Term Memory (LSTM) network, a recurrent neural architecture designed to learn patterns and dependencies in sequential time-series data, and it is used because it can model temporal structure in engineered financial features across fixed-length historical windows. Model uncertainty is summarized using holdout residual variability to provide a confidence interval for predicted prices. The trained model, scaler, and metadata deployed through an interactive Streamlit application that displays results including predicted returns, uncertainty range, sentiment summary, and optional backtest comparisons.

Hua Hua is an autobiographical fiction graphic novel featuring seventeen stories across twenty-one pages. Across these stories, the graphic novel focuses on the key memories of Mei, a young Taiwanese-American woman, as it follows her chronologically from early childhood to young adulthood. It focuses on her challenges growing up as an outcast at home and in school, and her constantly evolving sense of self. Throughout, she questions and evaluates her relationship with her culture, family, friends, and former acquaintances. Depending on the audience, Mei learns which aspects of herself she should hide away and others she can safely express. Her early childhood focuses on the struggles of growing up as the only Taiwanese student in her predominantly white elementary school. It explores her earliest encounters with self loathing from the constant teasing and badgering of racial and cultural differences. Meanwhile, her teenage years focus on self discovery, retrospection, and acknowledging generational trauma and destructive behavior stemming from her relationships. Meiís story culminates in the acceptance of the experiences, positive and negative, that made her the sum of who she is. With these stories, the graphic novel aims to balance the comedy and troubles that come with growing up in an environment that ostracizes those who are unwilling to conform. It conveys that bitter experiences in the moment can be looked back on as through the lens of a story with a lesson or punchline.

Inflammatory Bowel Disease (IBD) is characterized by inflammation along the gastrointestinal tract that causes diarrhea, cramping, weight loss, and fatigue. IBD has been associated with the loss of function of the Protein Tyrosine Phosphatase non-receptor Type 2 (Ptpn2) gene. Our group has reported that whole-body Ptpn2-knockout (KO) mice have decreased numbers and functionality of specialized small intestinal epithelial cells called Paneth cells, increased gut barrier permeability, and a dysregulated gut microbiome, all of which are factors in IBD development. Intestinal epithelial cells (IECs) directly maintain gut microbiome homeostasis, control gut permeability, create the mucus layer, and absorb nutrients, thus Ptpn2 deletion could directly influence multiple IEC functions. Moreover, whole-body Ptpn2 deletion causes a hyperactive immune response in mice, which significantly compromises the differentiation and function of IECs. Here, we investigated whether intrinsic epithelial Ptpn2 loss affects the function and differentiation of IEC subtypes in vitro and in vivo. Methods: Enteroids were generated from small intestinal crypts isolated from Ptpn2-wildtype (WT) and global Ptpn2-KO mice (BALB/c). In addition, tamoxifen-inducible epithelial-specific Ptpn2-KO mice (Ptpn2?IEC) and Ptpn2 controls (Ptpn2fl/fl; C57Bl/6J) had IECs isolated for total RNA isolation followed by qPCR analysis of gene targets associated with function and differentiation of IEC subtypes. Results: In undifferentiated enteroids, the expression of IEC subtype-associated genes Lyz1, Defa5, and Defa6 (Paneth cells), Muc2 and Tff3 (goblet cells), c-Maf (enterocytes), and Sucnr1 and Pou2f3 (tuft cells), showed no significant difference in Ptpn2-WT versus KO enteroids. In addition, there was no change in the expression of the same targets in ileal IECs of Ptpn2?IEC versus Ptpn2fl/fl mice (n>5), except for Paneth cell markers that were all downregulated as reported by the McCole lab in a previous publication. However, immunostaining for Dclk1, a tuft cell marker, revealed higher numbers of tuft cells in the ileum of Ptpn2?IEC vs. Ptpn2fl/fl mice (n=6), while alcian blue staining revealed elevated number of goblet cells in the colon of Ptpn2?IEC vs. controls (n=8). Conclusion: Epithelial Ptpn2 deletion does not impact the expression of IEC markers in vitro, but it does increase the number of tuft and goblet cells in vivo. This implies a requirement for non-epithelial cell types in epithelial Ptpn2 regulation of IEC differentiation.

This literature review examines the emergence of human-AI relationships as an extension of parasocial relationships, focusing on the psychological, cognitive, and ethical factors that shape these interactions. Parasocial relationships have traditionally been defined as one-sided emotional bonds formed with media figures (Horton & Wohl, 1956), but advances in artificial intelligence have introduced interactive systems that can simulate reciprocity and personalization. Drawing on research in social cognition, personality psychology, and human-computer interaction, this paper explores how AI systems engage mechanisms such as anthropomorphism, attachment, and emotional regulation. Particular attention is given to how AI mirrors usersí language, tone, and preferences, creating a sense of intimacy that may blur the boundary between tool and social entity. The review also considers how individuals use AI in both functional and relational contexts, as well as how corporate design strategies optimize engagement through personalization and emotional validation. Although AI may appear to offer guidance or companionship, it does not possess consciousness or independent decision making ability. Instead, it generates responses based on predictive modeling and reinforces multiple perspectives, leaving users to determine the outcome. This dynamic can contribute to emotional reliance while maintaining the illusion of agency. Human-AI relationships are conceptualized as an interactive evolution of parasocial bonds that intensify perceived intimacy through responsiveness. Understanding these processes is essential for evaluating the ethical implications of AI integration and promoting informed, critical engagement with relational technologies.

Deepfakes have become increasingly realistic with the rise of diffusion models, vision transformers, and multimodal generators, posing major risks to security, trust, and authenticity in digital media. While existing detectors achieve strong benchmark accuracy, their performance deteriorates under real-world distortions such as compression, transmission loss, or bitrate fluctuations. This literature review traces the evolution of deepfake detectionófrom early spatial and temporal CNNs to frequency-, latent-, and transformer-based architecturesóand examines robustness-oriented frameworks including ADD, QAD, BZNet, and DPL. To address the gap between laboratory conditions and deployment environments, this study introduces two complementary research directions. First, an adaptive detection framework, the MDN-Guided Noise-Aware Modular Ensemble, models test-time uncertainty using a Mixture Density Network to dynamically select noise-specific fine-tuning modules. Second, a new dataset, DeFND (Deepfake Forensics under Network Degradation), captures authentic Wi-Fi and cellular transmission artifacts by streaming deepfake videos through real WebRTC pipelines. Together, these contributions provide a foundation for evaluating and improving deepfake detectors under realistic network and corruption conditions, advancing the pursuit of reliable, uncertainty-aware detection systems for practical deployment.

In the digital age, social media has fundamentally altered the relationship between corporations and consumers, creating new mechanisms of public accountability. This paper examines the long-term financial effects on firms involved in major social media activism events, analyzing three case studies: Starbucks, Target, and Costco. Through qualitative analysis of public sentiment across TikTok, Instagram, and X, alongside quantitative financial data including stock price fluctuations and quarterly revenue reports, this paper identifies correlations between online consumer behavior and corporate financial performance over a one-year period following each event. Findings suggest that while social media activism produces measurable short-term financial consequences, its long-term impact is limited by activism fatigue, discourse displacement, and the resilience of large corporations. Furthermore, negative and positive sentiment appear to produce different financial outcomes, with negative activism more strongly impacting investor confidence and positive activism more significantly influencing consumer revenue. These findings begin to speculate the tangible intersection of digital consumer behavior and corporate financial accountability.

A brain aneurysm is a bulging artery inside the brain that can rupture, often causing severe debilitation or death. A potential patient-specific treatment involves using a polyurethane shape memory polymer (SMP)-based foam that can be compressed for endovascular surgical insertion and then expanded with heat to maximally fill the aneurysm space. By blocking blood flow into the aneurysm, the customizable SMPs reduce rupture risk and improve patient outcomes. However, their long-term mechanical durability and shape recovery capabilities must be well understood before they can be safely implemented in medical applications. The goal of this research is to examine the durability, shape retention, and recovery time of SMPs through repeated compression tests, so this material can eventually be used to treat patients. A total of 18 SMP foams were fabricated with 10%, 15%, and 20% infill densities using fixed ratios of three monomers: (i) hexamethylene diisocyanate (HDI), (ii) N,N,N0,N0-tetrakis (hydroxypropyl) ethylenediamine (HPED), and (iii) triethanolamine (TEA). Nine samples underwent 15 cycles of compression tests at 60 ∞C to characterize the materialís stress-strain behavior, with recovery time recorded by a camera, and microscope images taken every five cycles to assess any structural changes or damage. The remaining nine samples underwent 10 cycles of compression at 60 ∞C, followed by another 10 cycles of compression at room temperature (~20 ∞C). The stress-strain results were compared across infill densities and recovery levels to determine how repeated compression influences the SMPsí thermo-mechanical properties. This research showed that repeated compression cycles caused reduced stress, structural changes in overall morphology, and variations in recovery time. The SMP foam cubes were able to withstand the repeated compression cycles and retain their functionality, showing promise in their application in future medical applications.