All Capstone Projects
Pharmaceutical companies committing fraud (through illegal kickbacks, off-label promotion, or false claims submissions) leave detectable traces in their SEC financial disclosures years before legal action materializes. This project builds a complete, end-to-end pipeline that (1) automatically extracts domain knowledge from historical DOJ complaint filings using a large language model platform...
Auto-eHRmonize Flowsheet is a Python package designed to solve a critical interoperability problem in healthcare informatics: clinical flowsheet label harmonization. Healthcare facilities use heterogeneous naming conventions for identical clinical measurements (e.g., “Heart Rate”, “HR”, “Pulse”, “Heart rate BPM”). This tool automatically maps these varying labels to standardized, reference terminology using...
This work presents an evaluation framework for explainable and adversarially robust sleep stage classification using multivariate time series data collected from wrist-worn wearable devices (DREAMT dataset). We examine three models commonly used for sleep stage classification across industry and research: a LightGBM plus biLSTM for post-processing, a support vector machine...
FinCom (Financial Committee) is a supervisor-based multi-agent framework designed to support structured, committee-style financial decision-making. Built on LangGraph, the system orchestrates three domain-specialized agents—Research, Quantitative, and Risk Management—each powered by large language models with ReAct-based reasoning and constrained tool access. These agents collaborate under a central Supervisor to analyze financial...
Financial institutions generate regulatory compliance reports such as Suspicious Activity Reports (SARs), Currency Transaction Reports (CTRs), and sanctions filings under strict legal requirements. These workflows remain highly manual, requiring analysts to review transaction data, draft narratives, and complete structured reports, resulting in significant cost and inefficiency. On average, a single...
This analysis investigates whether the lexical structure of a dream-defined by how characters, locations, and actions interlink-serves as a biomarker for psychological health. Utilizing Structural Network Analysis, we aimed to identify if higher levels of mental distress correlate with either fragmented narrative networks or overly rigid, repetitive clustering patterns in...
Background Compass Coffee is a Washington, D.C.–based coffee roaster and café chain operating multiple locations across the region. They started in 2014, with the mission “to make peoples’ days better”. Unlike many other restaurants, all Compass’s ingredients are delivered fresh, including items other than coffee, such as bakeries and snacks....
This capstone project addresses a concrete decision-support problem in financial risk management: analysts face a high-volume stream of macroeconomic and market news, but only a small subset of articles contains actionable information for foreign -exchange risk hedging. The project’s final scope, as documented in the clean-up branch and the capstone...
This project develops a scalable ICU Digital Twin framework to model patient trajectories and predict clinical deterioration using electronic health record data from MIMIC-IV. ICU environments are characterized by rapidly evolving patient conditions, irregularly sampled data, and complex multi-modal signals. Those challenges render traditional models insufficient for both outcome prediction...
Although access to nutrition is a critical determinant of academic success, participation in subsidized school meal programs varies considerably among eligible schools. The Community Eligibility Provision (CEP) is a federal policy that allows schools and districts serving high-poverty populations to provide meals to all students at no cost. We estimate...
The Conservation Fund (TCF) manages large timberland properties where accurate forest stand boundaries and stand-level attributes are critical for inventory, management, valuation, and long-term planning. However, much of this work is currently performed manually, making it time-intensive, costly, and difficult to scale consistently across properties. This project addresses that challenge...
This Capstone Project develops a solution for banks to monitor their exposure to climate risk impacts on the U.S. property and casualty (P&C) insurance market, combining national-scale predictive modeling with a geographically granular Bayesian forecasting framework for California. Project Goal and Context The core problem is that increasing frequency and...
This capstone project develops an end-to-end Marketing Mix Modeling (MMM) and optimization framework for Citizens Financial Group to improve the allocation efficiency of marketing spend across multiple marketing channels and products. Traditional attribution systems frequently over-credit last-touch digital interactions while underestimating the cumulative impact of offline and upper-funnel channels. To...
Becoming rentABLE aims to help travelers with disabilities find truly accessible short-term rentals. The platform currently verifies accessibility features manually by reviewing listing photos, a process that is slow, subjective, and difficult to scale as the number of listings grows. Hosts often upload inconsistent or incomplete photos and make verification...
This project builds a pipeline that finds what IT company executives are talking about on their earnings calls, and flags when something new shows up. We pulled transcripts from 30 IT companies between 2023 and 2025, kept only the executive speech, and broke it down to the sentence level. Each...
Our project focuses on enhancing the safety and cultural robustness of text-to-image (T2I) models by addressing gaps in existing evaluation datasets. We began by contributing to improving election-related safety at Meta, enhancing AI safeguards against political misinformation and strengthening the accuracy of classifiers. In the process, we identified a lack...
ChessMindAI is a platform focused on improving chess education by combining advanced artificial intelligence with classic gameplay. Our current project explores how AI can assist in writing expert-level commentary for a chess textbook by generating insightful explanations for individual board positions. To achieve this, we use a custom, agentic workflow...
Buildings account for approximately 9% of global greenhouse gas emissions each year, yet emissions data for the building sector is often outdated or too low-resolution to support effective policymaking. Climate TRACE—a global nonprofit coalition that independently monitors emissions across 83% of global sources—seeks to fill this gap. As part of...
Conventional object detection models like YOLO often struggle with occluded objects due to reliance on single viewpoints and limited scene context. In our capstone project, we benchmarked BEVFormer, a transformer-based model which using bird’s-eye view representation, against YOLO on the NuScenes dataset. The NuScenes dataset contains diverse urban scenes with...
This project is an enterprise-grade log monitoring and anomaly detection system developed to support Data Oceans and their clients in identifying unusual patterns across large-scale transactional and system log data. The solution leverages a combination of unsupervised machine learning models (such as Isolation Forest and K-means) and statistical methods to...
In our capstone project we implemented unsupervised machine learning clustering methods to segment users of the real-estate website Monopolio to provide insights for marketing purposes. Our client, DD360, has a new website that presents real estate opportunities for rent and purchase in Mexico, where users can browse through different neighborhoods...
Critical care providers need to make high-stake clinical decisions on a daily basis. Making research-informed decisions is challenging because of the sheer amount of research results published in the field everyday with potentially conflicting conclusions regarding the same clinical questions. This project aims to develop a medical question answering model,...
Duke Health’s Anesthesiology team faced a significant challenge in analyzing vast amounts of patient-related data due to the limited technical skills of their team members. This limitation forced them to rely heavily on scarce statisticians and programmers for data cleaning and processing, a process that was both time-consuming and expensive,...
The Federal Open Market Committee (FOMC) sets the target for the federal funds rate, a critical tool influencing interest rates, inflation, and the broader economy. What if we could use large language models (LLMs) to anticipate their next decision? In partnership with Bank of New York (BNY), our project developed...
This project analyzes customer satisfaction at Citizens Bank using Net Promoter Score (NPS) survey data, focusing specifically on customers’ likelihood to recommend the bank to others. By combining survey responses with customer account data, transaction history, and banking behaviors, we identified key factors that drive customer satisfaction and loyalty. Our...
Facebook’s analytics currently provide reach metrics for fixed intervals (1-day, 7-day, and 28-day periods), limiting the ability to track reach progression over custom timeframes. To address this gap, our project develops an XGBoost model to estimate reach for any arbitrary number of days, incorporating transformations to handle non-linearity in the...
This project develops a multi-stage framework to support Duke Women’s Soccer recruitment through advanced modeling and a web-service platform. In the first-stage of modeling, detailed game-level player statistics from the Wyscout API are transformed through extensive aggregation and feature selection to form the final set of average, percentage, and versatility...
Our client, Socialinsider, provides social media analytics tools in measuring and comparing social media performances for business. We help our client by improving their marketing strategy to better convert subscribed users and expanding their product by predicting future social media performance metrics. To improve the conversion rate of subscribed users,...
Subsalt is experiencing high computational costs from generating legally de-identified synthetic data by training multiple generative models across various configurations. Many of these configurations are not utilized, contributing to the inefficiencies. These inefficiencies lead to unnecessary expenses, especially from running costly privacy tests on different models and configurations. To tackle...
Let’s make vacation accessible for everyone–that’s the mission of Becoming rentABLE, and it becomes ours too. Our objective focused on helping Becoming rentABLE to expand their inventory of accessible short-term rental rental across 18 target U.S cities. To achieve this, we analyzed Airbnb text review data and used a RoBERTa-based...
As a strategic data initiative with AB InBev, our team was tasked with exploring the use of alternative data to enhance their Business-to-Business (B2B) product recommendation systems. Traditionally, AB InBev relied heavily on internally generated data—such as customer purchase activity—for these predictions. We were brought in to investigate whether novel,...
Citizens Bank wanted to understand the impact of different marketing channels on sales while accounting for macroeconomic factors and regional presence. To address this, our team developed a Marketing Mix Model (MMM) that categorized banking products based on similar sales trends and built separate models for each category. The approach...
Our client, Orderly, aims to develop an up-to-date healthcare provider database to help patients, healthcare insurance firms and health tech companies to find the right healthcare provider information efficiently. Currently, they have developed an algorithm capable of verifying healthcare provider information by finding relevant online data. Our role is to...
Public project statement not available.
Our project with DataOceans represents a significant advancement in solving the intricate challenges of data preprocessing for key industries by leveraging artificial intelligence, including machine learning and natural language processing. Through the implementation of an AI-driven system that features error detection and machine learning for header classification, especially employing the...
The CR system matches student-athletes with suitable college sports programs by analyzing students’ characteristics and preferences which are then used to identify programs that best align with them. Our analysis shows that our recommendation system accurately predicts the school a student actually committed to 33% of the time, significantly outperforming...
Our team collaborated with AMECO, a construction service company, to address their challenge of aligning actual profit margins with targeted goals. Based on past years sales performance, AMECO is experiencing a constant decline in profit. The target profit is achieved by selling their products at their target profit margins. Therefore,...
This project characterizes the relationship between gender and academic research funding through the examination of the allocation of National Institutes of Health (NIH) R01 grants. Despite providing detailed funding and institution-level information on Principal Investigators (PIs), the NIH does not make individual-level demographic data public, making it challenging to directly...
In our Capstone project, we provided a pipeline for our clients to evaluate and reduce bias in their financial classification models using the AI Fairness 360 package developed by IBM. Our client, 2nd Order Solutions (2OS), is a financial service firm that provides banks with machine learning algorithms to decide...
Anesthesiology providers are believed to be at least as likely as the general population to suffer from non-medical opioid use. Currently, there is a lack of data-driven methods to efficiently identify providers in need of support. This project applies machine learning models to electronic health records to detect changes in...
Our client, DataOceans distributes other companies’ billing statements that contain advertisements for new products. To make the ads more personalized, DataOceans has developed a user interface that allows businesses to define specific criteria for targeting their advertisements. For instance, a business might target married males over the age of 40,...
All animal behavior is based on approach and avoidance motivations. Organisms tend to approach things that are positively valenced or beneficial to them and avoid things that are negatively valenced or harmful to them. The automaticity of these fundamental motivations has been supported extensively with empirical research using various types...
Our project addressed the significant challenge of reducing computational and storage costs associated with the deployment of large language models (LLMs) at Proofpoint. Recognizing the vital role of LLMs in enhancing the detection and mitigation of cyber threats across emails, documents, and other communication channels, our team embarked on a...
It is widely believed that environmental factors such as lack of employment opportunities, Inadequate access to food, unstable housing, and household substance use, lead to family stressors that negatively impact children. Some of these impacts may come through direct maltreatment from family or community member. This capstone project focused on...
The US Army is responsible for recruiting for its Special Operations Command, a unit of highly specialized personnel responsible for challenging and critical missions, where its members come from various branches in the larger military (Army, Navy, and Air Force). The recruiting and hiring process is very time-consuming, and the...
Machine learning models used in production settings are frequently retrained as new data becomes available to improve accuracy. However, this causes inconsistencies between the predictions provided by different versions of the model, a process known as prediction churn. This is undesirable for end users and organizations, as it might result...
Despite the availability of diverse geographical data, annotating it for analysis is time-consuming, requires specialist knowledge of the specific application domain, and existing supervised models based on domain-specific labeling have limited applicability. This expensive and time-consuming process results in unequal access to datasets, especially in low-income countries. To overcome these...
The use of Electronic Health Records (EHR) by medical students during their clinical rotations provides a rich source of documentation of the students’ clinical experiences. However, accessing and summarizing those notes through the EHR is cumbersome and inefficient. There is a need to improve the digitization of notes and encounter...
Discrepancies between population and housing estimates and the 2020 Census counts, especially in undercounted or overcounted areas, might lead to poor policy decisions. These disparities, often correlated with various demographic attributes, might increase inequities in resource distribution. To address this issue, we conducted a comparison of current population and housing...
Interpretability is a very key aspect of 2nd Order’s solution to provide financial institutions with explanations to customers for certain financial decisions, such as credit card application denials as mandated under US law. In this project, we evaluated and compared the two most popular explainable prediction algorithms commonly used in...
Showing 1-50 of 82 results