Bishal Lamsal, Prasanthi Sreekumari

Many people spend a large amount of time doomscrolling on platforms like LinkedIn and X, where unnecessary and distracting content often appears even when users only want to see meaningful and useful posts. This constant exposure reduces focus and productivity. The objective of this study is to develop a web browser extension called BloomScroll that uses artificial intelligence to analyze a post’s caption, image, and metadata in real time and filter out unwanted or low-value content. The extension allows users to choose or prompt the BloomScroll with the type of content they want to see, ensuring a more personalized and distraction-free feed. The system works by continuously analyzing and filtering content based on user preferences as they scroll. The expected outcome is a significant reduction in wasted time, improved focus, and a noticeable increase in daily productivity.

ABDULLAH AL MOAEL

Negation remains a persistent challenge for modern language models and often leads to reversed meanings or factual errors. In this work, we conduct a causal analysis of how GPT-2 Small internally processes negation and related linguistic polarity transformations. We examine hidden representations at both the layer and attention-head level using a self-curated dataset of 12,000 matched pairs of affirmative and negated sentences covering multiple linguistic templates and negation forms. To quantify this behavior, we introduce the Negation Effect Score (NES), a metric that measures the model’s sensitivity to distinguishing affirmative statements from their negations.

We apply two causal interventions to probe internal structure. In activation patching, internal activations from affirmative sentences are inserted into corresponding negated runs to observe semantic shifts. In ablation, selected attention heads are temporarily disabled to test causal necessity. Our results show that negation processing is not widespread but instead concentrated in a small number of mid-layer attention heads, primarily within layers 4 to 6. Ablating these components increases NES on in-domain data, indicating weaker negation sensitivity, while reintroducing cached affirmative activations further increases NES, confirming that these heads primarily carry affirmative polarity signals. These effects persist across multiple negation forms indicating a localized, interpretable polarity circuit. 

Maya Trutschl, Steven Conrad, Kimberley Hutchinson

Pressure ulcers are injuries to the skin and underlying tissue over bony prominences due to pressure or friction, ranging from pink spots to bone-deep lesions. They are a major health issue that affects more than 2.5 million in the US annually, causing a greater mortality rate than any single cancer, except lung cancer, and affect the quality of care in hospitals and long-term care facilities. Preventing them remains a labor-intensive challenge: from manual risk assessment charts to inspections of the patient's skin and regular patient repositioning. Prevention devices range from specialized beds to mattress overlays, but most are too expensive, cumbersome, or not applicable to the ICU.

To predict and mitigate ulcers, I developed a two-step solution that can be integrated into electronic health records systems and automatically predict ulcer risk, combined with a device for patient monitoring. I used a data set with 546,028 hospitalizations and 65,366 Intensive Care Unit (ICU) stays to model and automatically predict ulcer risk. I separated the ICU data and can predict ulcers within 24 hours of admission with more than 95% accuracy using only basic lab values, demographic data, and comorbidities. My system is a non-invasive device that automatically tracks patient positioning. I built three prototypes and ran a clinical study at the local hospital (ICU and ward) where patients were monitored by my device for more than 150 hours with greater than 99% accuracy.

suman yadav, Prasanthi Sreekumari

Large Language Models (LLMs) such as ChatGPT have improved access to information, but their use in structured academic learning remains limited due to fragmented workflows, repeated material uploads, lack of persistent context, and the absence of learning analytics or reliability assessment, which reduces their effectiveness for sustained study. The objective of this study is to design and present Lernix, an AI-powered intelligent learning platform that addresses these gaps by enabling persistent, document-aware learning combined with learner analytics and trust-based AI evaluation. The study is conducted through the design and implementation of a unified system that allows users to upload academic resources once, interact with them across sessions using contextual retrieval, track engagement metrics such as time spent per topic and assessment performance, and evaluate understanding through integrated MCQ-based assessments with data-driven analysis. Lernix incorporates a Trust AI module that evaluates the reliability of AI-generated responses using document-grounded retrieval, semantic similarity, and contradiction detection to ensure academic alignment and transparency. The key findings demonstrate that Lernix transforms passive LLM usage into an active, adaptive, and trustworthy learning ecosystem by improving continuity, providing personalized recommendations based on learner data, and enhancing the reliability of AI-assisted explanations, thereby supporting 

Augustine Nwafor, Ricky Wiggins, Stewart Greathouse, Giovanni Solitro, Marjan Trutschl, Urska Cvek

This study presents a low-cost wearable system that estimates three-dimensional (3D) knee joint kinematics and detects ACL injury risk in real time using dual inertial measurement units (IMUs). 

Two IMUs were mounted on the thigh and shank at 10-15 cm from the knee joint, each oriented with the x-axis, y-axis, and z-axis. Quaternion-based orientation data were processed by a microcontroller to compute relative knee motion following the International Society of Biomechanics (ISB) joint coordinate conventions: flexion/extension about the x-axis, valgus about the z-axis, and internal rotation about the y-axis. An adaptive calibration routine defined a neutral standing pose for each participant. Real-time ACL risk detection was triggered when flexion was 10-30° and knee valgus exceeded 5°. Processed joint angles and alerts were transmitted via BLE to a dashboard for live visualization. 

The system achieved stable joint angle estimation at a 100 Hz sampling rate, with wireless streaming at 20 Hz. Preliminary field trials demonstrated consistent detection of high-risk valgus postures during squat and jump-landing tests, with an average signal noise below 2° after smoothing. 

This project simulates aspects of laboratory-grade 3D knee motion tracking using quaternion fusion and ISB joint mapping. Its portability, affordability, and real time feedback capabilities make it suitable for ACL injury risk assessment and rehabilitation monitoring.