Deepfake videos are becoming increasingly difficult to distinguish from real videos, leading to increased digital deception. Haptic feedback is often used to enhance realism in gaming and virtual interfaces, but its impact on deepfake videos is less clear. To understand how vibrotactile feedback influences video realism, a user study (n=4) was conducted with a custom interface that allowed participants to view and rate videos based on realism (0 = not realistic, 1 = very realistic) while receiving haptic feedback via a finger-mounted linear resonant actuator. A dataset of real and fake videos with haptic profiles was developed and used for this study. Preliminary pilot study results indicate that participants can differentiate real from fake videos (p < 0.001), but there was no significant effect of vibration on video realism (p = 0.43). In the future, the study will be extended with additional subjects, stimulus combinations, and types of haptic feedback modalities. Given the growing prevalence of deceptive digital practices, this project aims to identify factors that could alter the perceived realism of media.

While microservices are a popular architecture for large-scale applications, managing systems with hundreds of connected components makes it difficult to maintain a holistic understanding of the landscape. This complexity often slows development velocity and accumulates technical debt. To address this, we propose LumeNest - a tool that automates the architectural reconstruction of microservice communications through static analysis. LumeNest utilizes CodeQL for interprocedural dataflow analysis and subsequently produces a component connector diagram. By validating this method against a benchmark microservice system, we demonstrate that our pipeline successfully resolves cross-service endpoints without requiring code execution. Our approach clarifies complex system logic, providing a scalable means to document system logic and accelerate developer onboarding in large distributed systems.

The global STEM doctoral landscape is defined by a widening divergence in talent production, where foreign output exceeds that of the United States. This disparity is intensified by a navigation gap coming from a lack of pathways from classroom concepts to active research. Traditional recommender systems often suffer from pedagogical blindness, relying on popularity rather than logical requisites. This research background presents an experimental framework for automated cartography to construct educational knowledge graphs (EKG). We hypothesize that using LLMs to perform bottom up ontology discovery from unstructured curriculum data can bridge this navigation failure. Our methodology evaluates a two layer extraction pipeline. The first layer extracts core concepts from multimodal, unstructured data, while the second layer performs entity alignment to map instructional terms to research areas. We employ the CoDe-KG framework to model sentence complexity, converting complex sentences into simple forms. Accuracy is measured against human annotated standards using the Path Coherence Score (Cp). Preliminary findings indicate that syntactic decomposition reduces extraction errors to approximately 10%. This project is significant because it transforms static catalogs into agentic world models, providing students with a transparent critical path to the research frontier and building an educational system that advances the future of academic innovation globally.

Monocular 3D object understanding has largely been cast as a 2D RoI-to-3D box lifting problem. However, emerging downstream applications require image-plane geometry (e.g., projected 3D box corners) which cannot be easily obtained without known intrinsics, a problem for object detection in the wild. We introduce \textbf{MoCA3D}, a \textbf{Mo}nocular, \textbf{C}lass-\textbf{A}gnostic \textbf{3D} model that predicts projected 3D bounding box corners and per-corner depths without requiring camera intrinsics at inference time. MoCA3D formulates pixel-space localization and depth assignment as dense prediction via corner heatmaps and depth maps. To evaluate image-plane geometric fidelity, we propose \textbf{Pixel-Aligned Geometry (PAG)}, which directly measures image-plane corner and depth consistency. Extensive experiments demonstrate that MoCA3D achieves state-of-the-art performance, improving image-plane corner PAG by 22.8\% while remaining comparable on 3D IoU, using up to $57\times$ fewer trainable parameters. Finally, we apply MoCA3D to downstream tasks which were previously impractical under unknown intrinsics, highlighting its utility beyond standard baseline models.

Extracellular vesicles (EVs) are critical mediators of intercellular communication, but profiling secretion at single-cell resolution remains challenging. To address this, we developed a high-throughput microfluidic platform using nanovials (nanoliter-volume hydrogel microparticles with functionalizable cavities) to capture single-cell EV secretion profiles. We optimized scalable cell loading by fabricating 35 μm and 65 μm nanovials, tuning reagent flow rates for ideal dimensions and cavity morphology. Nanovial cavities were functionalized with anti-CD63 antibodies via a biotin-streptavidin bridge, then loaded with human induced pluripotent stem cells (hiPSCs) or induced mesenchymal stem cells (iMSCs). To isolate microenvironments and prevent cross-talk, nanovials were sealed with solid hydrogel capping particles, allowing locally secreted EVs to be captured by CD63. Captured EVs were fluorescently labeled with anti-CD9 and anti-CD81 antibodies. Flow cytometry was utilized to analyze and sort distinct single-cell EV secretion phenotypes, identifying CD9-dominant, CD81-dominant, and dual-positive states. This platform enables robust multiplexed profiling of EV heterogeneity in stem cell populations, offering a powerful tool for cell biology and translational medicine. Future work will adapt this platform for two-cell assays to measure EV-mediated outcomes on co-encapsulated target cells, enabling cell selection based on functional EV profiling rather than mere abundance.

Magneto-fluorescent proteins offer a promising strategy for generating externally controlled optical contrast, with potential applications in imaging and sensing in scattering tissue environments. This project focused on MagLOV2, a protein whose fluorescence changes under applied magnetic stimulation, and aimed to determine whether this response could be reproducibly measured and described using a minimal kinetic model. The study involved plasmid transformation and colony preparation in E. coli, wide-field epifluorescence imaging of fluorescent colonies under fixed illumination and square-wave magnetic stimulation, and computational analysis in MATLAB using a three-state ordinary differential equation model. Under an applied field of approximately 50 mT at 0.25 Hz, colony fluorescence showed repeatable decreases during magnet-ON intervals and recovery during magnet-OFF intervals. Separate photobleaching controls indicated that bleaching occurred on a slower timescale than the magnetic switching response. A minimal model consisting of fluorescent, dark, and photo-bleached populations captured the overall structure of the measured dynamics and provided an initial mechanistic framework for interpreting the response. Together, these results establish an experimental and computational foundation for studying magnetically modulated fluorescence and support future work on more realistic spatial models, comparative protein studies, and imaging-oriented simulations.