Company Relationship Prediction System Using Network Data
Team Members: Alex Li, Yiheng Yao, Grayson Gong, Frank Peng
Rel8ed wanted a system to predict whether two companies have a business relationship and, if so, identify the type of relationship. The goal is to simplify complex company networks into understandable patterns that can support decision-making and data analysis. To achieve this, we built a three-step prediction system. First, a logistic regression model determines whether a relationship exists between two companies. Second, a classification model (Extreme Gradient Boosting) predicts the general type of relationship. Finally, for cases involving ownership structures, a specialized model distinguishes between direct and indirect (ultimate)control using network-based features. The system achieved great accuracy and balance between precision and recall across different relationship types. Finally, the team delivered the detection system for Rel8ed that can identify both common and rare relationships, and handle real-world corporate network data effectively.
Lightweight Heat Stress Detection for Cow Health Monitoring
Team Members: Aakanksha Dutta, Duole Hong, Elvis Imamura, Sreejato Chatterjee, Eva Salomone
Cattle health monitoring is essential for maintaining productivity, but traditional methods rely on manual observation that often identifies illness only after it has progressed. This project aims to improve Zalliant’s temperature-based alert system by evaluating whether data-driven methods, including machine learning, can detect health issues earlier while remaining cost-effective. We reconstructed each cow’s normal temperature pattern by removing temporary drops caused by drinking and smoothing the data to create a stable baseline. We then applied a context-aware method that detects unusual increases in temperature based on how much they differ from normal behavior and flags only those that are significant and consistent over time. After testing machine learning approaches, we found they relied on limited labeled data and produced inconsistent results, making them less reliable. The final model performed well, identifying key temperature increases while limiting false alarms. It detected issues about 9 hours earlier than current methods and remained effective with less frequent data, helping reduce costs and improve early response.
Predicting What Drives Engagement in Mental Health Resilience Tweets
Team Members: Rusha Bhat, Sana Singh, Josh Feld, Omer Aktas, Vita Gu
University of Rochester Medical Center’s Clinical Translational Sciences Institute wants to better understand the promotion of mental health resilience on Twitter. Thus, the goal of this project is to identify which types of tweets are most likely to drive engagement in mental health resilience discussions, using a dataset of 100,000 tweets. The team implemented an end-to-end analytical pipeline combining manual annotation, large language model predictions, and statistical modeling. A subset of tweets was double-coded for relevance, sentiment, and key content features to ensure consistency: these labels were used to guide a model in classifying the full dataset. Regression was then applied to measure the impact of each feature on engagement outcomes. The results show that engagement is primarily driven by content characteristics.
Storytelling, clear structure, and positive, natural language significantly increase engagement, while slang and direct calls to action reduce it. This provided a consistent framework for understanding effective mental health communication online.
Portfolio Optimization Using Federal Open Market Committee Minutes
Team Members: Mia Alex, Yaxun Chen, Carter Schmitt, Tianyou Tu, Bryce Tyler
The Federal Open Market Committee (FOMC) is tasked with setting monetary policy in the United States to manage inflation and employment. They meet eight times a year and shortly after, release the minutes of their discussions to the public. The policies they set have broad-reaching impacts on the greater US economy. This project seeks to determine whether these FOMC minutes can help with stock predictions and portfolio optimization.
Three sets of data were collected over a period of around ten years. FOMC minutes were scraped from the Federal Reserve website. Large language models and traditional NLP techniques were used to create features from the minutes. Quantitative data on twenty different stocks and a set of macroeconomic indicators were sourced from the Bloomberg Terminal.
A two-step approach was developed to determine the usefulness of FOMC minutes. A stock prediction model (XGBoost), incorporating FOMC features, was created and compared against a model that did not use FOMC-derived features. The predictions of these models were then used to optimize two portfolios to determine which one would have a higher return.
Directional accuracy, which measures how often a model correctly predicts whether a stock’s price will rise or fall, rose from 54.9% to 56.46% when FOMC features were added. Two FOMC features were identified as significant contributors to model performance: one based on Latent Dirichlet Allocation and another representing current policy stance, generated using an LLM. Portfolio returns using FOMC-derived predictions also increased from 28.79% to 33.58%.
The results of our work show that incorporating features engineered from FOMC minutes can help with stock prediction and portfolio optimization. Future work could delve deeper into the usage of FOMC features. Our feature importance results only capture how individual FOMC features contributed to return predictions, and not any interactions between features. Additionally, different baskets of FOMC features could be experimented with.
Search Your Video: Turning Footage into Queryable Data
Team Members: Victoria Ritter, Audrey Clayton, Britton Ambrose, James Kaprowski
Video is one of the most information rich media, yet it is perhaps the most difficult to search through. Unlike text, images, audio, and animations that are stored as indexed records for fast searching, video is a stream of raw pixel or frame data for playback. Searching full length dance videos for words like “twelve chickens”, or attributes like “2 people”, or “high energy” is challenging. To address this, we developed a method to process full length dance videos into millions of short indexable video clips. We present a technique that segments each video into a time window (on the order of a second) and for each window applies OCR for transcribed speech, motion tracking of dancers, and activity measurement. Each window is tagged with simple tags like “solo”, “duo”, “group”, “high energy”, etc. We applied our method to a small data set of 6-8 videos for a total of over 20 minutes of video, and obtained over 100 window clips, indexed for cross-video searching using transcript text, tags, and motion features. This paper shows video becomes much more useful organized by language, movement, and time.
Initial Analysis of Monroe County Crashes, 2012-2024
Fernando Lopez, Justin Li, Cloud Xiao, Songtao Zhang, Barry Magenya
We were tasked by Reconnect Rochester with analyzing a database of Monroe County Crashes from 2012-2024, derived from police reports, with a special focus on 1) vulnerable road users (pedestrians and bikers) and 2) the effects of: Time of Day, Jurisdiction, Frequent Crash Corridors, and Road Projects (particularly: the Lake Avenue M&R, the East Avenue and Elmwood Avenue Road Diets, and the Dewey & Driving Park Intersection Realignment). We utilized a variety of statistical methods (logistic regression, Wilcoxian, etc.) to examine each facet in turn. With regards to Time of Day, we found a period of time, from approximately 20:00 to 06:00 for vulnerable users, where there is a statistically significant increase in the odds of a crash resulting in a fatality or serious injury; with regards to Jurisdiction, we found that there is a statistically significant increase in likelihood of a crash having a fatal or serious-injury on roads administered by the state – approximately 1.05x, relative to non-vulnerable users on a city road. With regards to Road Projects, we did not find statistically significant increases or decreases for total crashes or vulnerable user FSI for Lake Avenue, Dewey & Driving Park, or Elmwood Avenue. We did find statistically significant decreases in total crashes from the East Avenue Road Diet, but this is very likely confounded with COVID-19.
We would like to emphasize the preliminary and exploratory nature of our findings, which do not account for key factors such as COVID-19, the demographics of a particular crash area, or changes in categorical definitions over time.
Teaching the Teachers: AI-Powered Syllabus Builder Using Retrieval Augmented Generation
Team Members: Caelan Clayton, Veena Laks, Aabha Pandit, Ramneek Nandha
Akkadian is building an artificial intelligence platform to help instructors design stronger course syllabi, especially because many instructors don’t have formal training in evidence-based teaching practices. The goal of this project was to develop a Retrieval-Augmented Generation (RAG) system that gives instructors grounded feedback on syllabus sections and answers teaching-related questions using trusted educational resources and university policy documents.
To support this goal, the team built a structured knowledge base, developed a retrieval pipeline with metadata tagging, query expansion, source balancing, and deduplication, and created prompt designs that guide the model to use only retrieved evidence. The team also delivered a working chatbot interface that allows users to select syllabus sections and subject areas for more targeted feedback.
The final system achieved strong evaluation scores, including 0.82 recall, 0.98 relevance, and 0.97 faithfulness, showing that it can retrieve key teaching concepts and generate useful, supported responses.
Otolaryngology Surgical Resident Performance and Experience Level Prediction
Team Members: Nicole Cassara, Jayla Hall, Marie Kinchington, Soumya Konar, Calvin Tian
URMC-Otolaryngology wants to improve surgical resident feedback by determining what technical skills are most associated with experience level and Objective Structured Assessment of Technical Skills (OSATS) scores. The goal of this project is to predict experience level, quantified by the number of surgeries a given resident has performed, and OSATS scores using motion tracking data collected from surgical tools. The team employed a Support Vector Machine (SVM) binary classification algorithm to predict experience level and a Gradient Boosting Regressor algorithm for performance prediction (OSATS score). The team also explored some unsupervised modeling techniques, performing Principal Component Analysis (PCA) and k-means clustering to measure the similarity of resident and attending motion data. Ultimately, the models obtained relatively high accuracy and the team informed the sponsor of the motion features with the highest predictive power for each of the target variables.
Title: Real-Time Voice Assistant for Sales Call Recommendations
Team Members: Lexiang Yang, Zhengyang Zhu, Shixin Lin, Baichuan Duan
This project built a live voice assistant that listens to sales conversations and recommends relevant Paychex products during the call so representatives can respond faster and miss fewer opportunities. The goal was to turn spoken customer concerns into clear, useful suggestions in real time, while keeping the experience simple and non-disruptive for both the representative and customer. To do this, we combined live speech transcription, a rule-based signal detector for common customer pain points (such as pricing, service, compliance, and manual work), and a search-and-generation recommendation engine grounded in Paychex product documents. We also added a verification step before showing recommendations and linked outputs to official product pages to improve trust and reduce errors. In testing and demos, the system produced recommendation cards in about 3 to 5 seconds, supported decisions using a knowledge base of 4,713 document chunks across 187 offerings, and showed stronger trigger reliability and recommendation precision over the earlier baseline.
AI-Powered Daily Portfolio Risk Report Generator from News Event Feeds
Team Members: Sylvia Liu, Yi Ren, Jiamu Tang, Youjia Zhang
Portfolio managers must review hundreds of news articles daily to identify risks across their investments, a process that is slow, inconsistent, and difficult to scale. Sponsored by FinchAI, this project developed an automated system that converts daily news event feeds into structured portfolio risk reports that can be reviewed in under ten minutes.
The team built a multi-stage pipeline using large language models (LLMs) to classify company news into one or more of seven business risk categories, then used deterministic aggregation logic to produce company-level severity ratings, preserve supporting evidence, and generate a narrative report.
The system was evaluated on 165 manually labeled articles across 78 companies, achieving approximately 90% label-wise agreement with human reviewers, with the strongest performance in Supply Chain and Product, People and Management, and Legal and Regulatory risks. The final deliverable is an interactive web application that allows users to select a portfolio, generate a risk report, explore financial analytics, and query holdings in natural language.