LLE Dynamic Knowledge Graph
Team Members: Wanting Shan, DongQi Lyu, Kangcheng Zhao, Shibo Xu Tulga, Shakleen Ishfar, Yiyao Tao, Yuxin Sa
The goal of this project was to help the University of Rochester’s Laboratory for Laser Energetics (LLE) navigate large collections of scientific literature by building an interactive knowledge graph that reveals meaningful relationships among research papers. The team processed and standardized a large metadata corpus, generated concise paper summaries using the BART-Large-CNN deep learning model, and represented papers with semantic embeddings. Using cosine similarity, the system compares user queries against all papers to identify the most relevant results and visualizes them in a hierarchical, tree-based knowledge graph. Users can explore nodes to access detailed metadata, abstracts, and summaries. The project also explored logic-chain connections between papers to evaluate the stability and coherence of the knowledge graph when individual nodes are removed. The final system enables faster discovery of related research, highlights hidden connections, and supports efficient exploration of complex scientific domains.
Fast Laser Predictions using Neural Networks to Support Real-Time Experiment Adjustments
Team Members: Nishant Hegde, Tim Yee, Jiamin Yin
The Laboratory for Laser Energetics aims to enable rapid adjustments between laser shots by accurately predicting laser behavior before the next shot is fired. This project focused on predicting ultraviolet laser beam characteristics over time using infrared measurements across multiple energy levels. The team applied deep generative neural network architectures, including the U-Shaped Fourier Neural Operator (U-FNO) and Attention Recurrent Residual U-Nets (Attention R2-UNet), to model spatial beam shape, pulse shape, and energy behavior. Training strategies incorporated specialized loss functions and memory-efficient techniques such as reduced-precision arithmetic to handle large, high-resolution datasets. The final models achieved low prediction errors (approximately 1%) in both spatial and temporal domains, demonstrating strong potential for supporting real-time laser experiment adjustments.
Emerald Ash Borer Outage Risk Prediction for Circuit Spans
Team Members: Prerna Ravinder, Harris Mazhar, Daniel Maass, Alex Korshunov
This project aimed to help Avangrid reduce power outages caused by ash trees weakened by the Emerald Ash Borer (EAB), an invasive insect that increases the risk of trees falling onto power lines. The team developed a predictive framework to identify which overhead power line spans are most likely to experience EAB-related outages. Their approach combined spatial analysis, statistical testing, and a custom voting-based prediction model that incorporated proximity to ash trees and surrounding vegetation density. The final model successfully identified a majority of outage-prone spans while maintaining reasonable precision, offering an interpretable and actionable tool for proactive vegetation management. Additionally, the team developed a Power BI dashboard to visualize high-risk power lines and support outage-prevention planning.
Evaluation of Cross-Sensitivity Alert (CSA) Suppression on Beta-Lactam Utilization
Team Members: Elijah Soh, Ludia Pack, Claire Kim
The goal of this project was to evaluate how removing beta-lactam cross-sensitivity alerts from the electronic health record (EHR) affected prescribing behavior and patient safety at the University of Rochester Medical Center. While these alerts were originally intended to prevent allergic reactions, emerging evidence suggests that most penicillin allergy labels are inaccurate and that true cross-reactivity among beta-lactam antibiotics is low. The team analyzed over 6,000 allergy-related alert encounters using pre- and post-intervention data. Methods included exploratory data analysis, stratified comparisons across clinician groups and shifts, alert-fatigue analysis using timing-based metrics, and logistic regression to assess subgroup-specific changes in beta-lactam use. Results showed increased beta-lactam utilization following alert removal, with no observed increase in adverse safety outcomes such as allergic reactions or anaphylaxis. Overall, the findings support improved antibiotic utilization without added patient risk.
Predicting Gluten-Free Customers among Wegmans Shoppers
Team Members: Reshmii Bondili, Snigdha Shrivastav, Priyanshu Rawat, Yuan Yao
Wegmans sought to improve engagement with gluten-free customers by moving from purchase-history targeting to a model-driven approach. The goal of this project was to identify customers who follow a gluten-free lifestyle based on their overall shopping behavior. The team integrated customer survey data with one year of sales, product, and store data and engineered features capturing gluten-free purchasing patterns, department-level engagement, and lifestyle indicators. Using XGBoost with regularization and class-imbalance handling, along with interaction features and principal component analysis (PCA), the team developed predictive models to classify gluten-free customers. A two-model approach performed best: one model broadly identified gluten-free shoppers, while a second more precisely identified customers with celiac disease. The solution provides Wegmans with a scalable, data-driven foundation for targeted engagement and performance measurement.
CellScapeXR: Virtual Reality Explorer for Spatial Gene Expression and Tissue Structure
Team Members: Evan Platten, Lydia Levesque, Mariah Nuzzo, Michael Seluanov
CellScapeXR is a virtual reality application designed to allow researchers and students to explore spatial gene expression and tissue structure in an immersive 3D environment. The project processed Visium spatial transcriptomics data using open-source Python tools, including SCANPY and Squidpy, applying dimensionality-reduction techniques (PCA, UMAP), clustering methods (Leiden), and normalization and batch-correction procedures. Processed results were integrated into a Unity-based 3D scene using custom C# scripts, enabling each data point to be linked to its tissue location and metadata. Additional analysis with STdeconvolve and Enrichr supported cell-type inference and heightmap generation to visualize layered tissue structures. The final prototype allows users to navigate UMAP point clouds, map points back to tissue locations, and explore cell-type distributions, demonstrating how immersive VR can make high-dimensional omics data more interpretable and accessible.
Visual Analytics for Patrol Assignment Optimization
Team Members: Don Nguyen, Srujana Chintala, Sakshi Hegde
The University of Rochester Department of Public Safety (DPS) requires improved visibility into incident trends, officer assignments, and resource utilization. This project developed a business intelligence platform that transforms raw incident data into actionable insights through automated pipelines and interactive dashboards. The team built a Microsoft Fabric lakehouse using a medallion architecture (Bronze, Silver, Gold layers) to process over 262,000 historical incidents. A semantic layer with pre-aggregated metrics supported five Power BI dashboards tailored to different stakeholders, including executive, operational, geographic, temporal, and optimization views. The team also developed a Mixed Integer Linear Programming framework to demonstrate how optimized officer assignments could improve incident matching rates. The platform enables data-driven decision-making for operational planning and resource allocation.
Identifying & Predicting Cognitive Trajectories in Early Parkinson’s Disease
Team Members: Linglan Yang, Ivy Wen, Jason Wang, Yarong Xiao
This project aimed to identify clinical, cognitive, and demographic factors associated with cognitive decline in early-stage Parkinson’s disease (PD) and to build predictive models for cognitive trajectories. Using longitudinal data from the Parkinson’s Progression Markers Initiative (PPMI), the team analyzed patients diagnosed within three years of baseline and followed for five years. Cognitive status was assessed using MoCA and NP1COG scores. Exploratory analysis identified age, olfaction (UPSIT), and non-motor burden (GDS) as strong predictors of decline. Blood biomarker analysis highlighted urea nitrogen (BUN) and eosinophils as statistically significant indicators, with BUN showing consistent associations with worse cognitive outcomes. An XGBoost classifier achieved a macro-F1 score of 0.532 for multi-class prediction and approximately 76% accuracy for binary classification. Results indicate that baseline clinical and metabolic measures provide early signals of long-term cognitive decline in PD.
Map to ROI: Hierarchical Media Mix Modeling — Understanding Regional Differences in Advertising Effectiveness
Team Members: Anirudh Vishwanath, Kiran Raj Paramasivam, Hariprakash Karthikeyan, Pournami Prakash
This project helped Butler/Till understand how advertising effectiveness varies across regions in the United States. The team combined national and regional datasets tracking weekly advertising activity and revenue, organizing the data by market to identify regional patterns. A Bayesian hierarchical media mix model was applied to capture delayed advertising effects while allowing information sharing across similar regions. This approach enabled smaller markets with limited data to be included in the analysis. Results showed that media channels perform differently across regions, with certain advertising types more strongly associated with sales in specific markets. The project provides a data-driven foundation for comparing regional performance and informing future media investment strategies.
Central Bank Communication Sentiment Analysis Platform
Team Members: Murat Al, Praveen K. Anwla, Saruultug Batbayar, Yibin Wang
Central bank policy statements and speeches significantly influence financial markets but are often complex and difficult to interpret. This project developed an automated platform to analyze central bank communications and summarize whether policy signals indicate tightening or easing economic conditions. The team applied natural language processing techniques, including a fine-tuned text classification model to score policy tone and a large language model to generate clear, human-readable explanations. The system collects documents from the Bank of England and the Bank of Canada, organizes them into monthly summaries, and reuses processed data to ensure fast response times. Interactive visualizations and written explanations are delivered through a user-friendly dashboard, enabling users to explore policy trends over time and improving accessibility for non-expert audiences.