Capstone Posters#
This gallery collects some of the completed capstone posters and short project descriptions drawn from the final documentation executive summaries.
Banking
Decoding the Dollar: Real-Time FX Forecasting with LLMs
The Bank of New York Mellon Corporation
This project supports financial risk analysts by filtering high-volume macroeconomic and market news for information relevant to foreign-exchange hedging. The pipeline ingests news, extracts article metadata and content, enriches articles with macro context, classifies DXY-relevant events with a large language model, and evaluates whether critical classifications correspond to large market reactions.
View poster PDFFinance
ClearPath: Multi-Agent Investment Advisory Committee
Clear Path
FinCom is a supervisor-based multi-agent framework for structured financial decision-making. Built on LangGraph, it coordinates research, quantitative, and risk-management agents under a central supervisor so investment questions can be analyzed from multiple perspectives and synthesized into actionable recommendations.
View poster PDFNGO
Streamlining Accessibility Verification in Short-Term Rental Platforms
Becoming rentABLE
Becoming rentABLE helps travelers with disabilities find accessible short-term rentals, but verifying accessibility from listing photos is slow and difficult to scale. This project develops a computer-vision review assistance system that highlights relevant accessibility features, organizes evidence at the property level, and supports human reviewers without replacing their judgment.
View poster PDFEducation
Does Access to School Meals Improve Student Outcomes?
Duke World Food Policy Center
This project estimates the effect of Community Eligibility Provision participation on academic outcomes for North Carolina students in grades 3-8. Using staggered difference-in-differences methods, the team finds modest but significant improvements in math and reading scores, with larger effects in lower-income districts and districts with lower prior academic performance.
View poster PDFHealth
Beyond the Black Box: Building Trustworthy AI for Wearable Sleep Trackers
Duke TRUST Lab
This project evaluates explainable and adversarially robust sleep-stage classification for wearable-device time series. The team compares LightGBM plus biLSTM post-processing, SVM, and CNN-RNN models, then studies the trade-offs among predictive performance, robustness to noise, and explainability for real-world sleep monitoring systems.
View poster PDFParmPharmaceutical and Compliance
AI-Driven Compliance Risk Intelligence for Life Sciences
GSK
This project builds an end-to-end pipeline for detecting compliance risk signals in life-sciences financial disclosures. It extracts domain knowledge from historical DOJ complaint filings, maps those insights to signals in SEC 10-K and 8-K filings, and trains a misconduct classifier using nested cross-validation so legal-risk patterns can be surfaced before enforcement actions appear.
View poster PDFEnvironment
Forest Stand Classification in the Southeastern United States
The Conservation Fund
The Conservation Fund manages timberland properties where accurate forest stand boundaries and attributes are essential for inventory, valuation, and long-term planning. This project develops a reproducible geospatial workflow that combines satellite imagery, disturbance history data, and the Segment Anything Model to automate stand delineation and generate stand-level attributes at scale.
View poster PDF