Audit Workflows, Reef Cameras, Tokenizer Research: Three CDS Capstone Projects
Most graduate-level coursework ends with a final exam or paper. The CDS Capstone Project ends with research the UN features on its blog, posters that win awards, and conference submissions to ACL. Two Fall 2025 student teams illustrate the range.
The Capstone Project is a course in the final year of the MS program that pairs students with industry partners and research labs to apply data science to real problems. Project areas include machine learning, natural language processing, statistical analysis, computer vision, and software engineering. Students get a problem that an outside partner cares about getting right, and a semester to deliver something useful.
One Fall 2025 team built a computer vision system to identify Caribbean reef fish from underwater rover footage. CDS MS students Amaan Mansuri, Vishwa Raval, and Shravan Khunti worked with the UNDP Accelerator Lab for Barbados and the Eastern Caribbean to develop a real-time pipeline on YOLOv11 that detects three species — surgeonfish, parrotfish, and grunts — under conditions where red light disappears past ten meters. The work was selected by the Japan Cabinet Office as part of the 2025 Japan SDGs Challenge and is featured on the UNDP Barbados blog. Their poster won Best Capstone Poster of 2025. The team’s most useful finding was unfashionable: a 9.4-million-parameter model outperformed one fourteen times larger, and ensembling five of their best models pushed accuracy past their 70 percent target. Data quality, not model size, was the binding constraint.
A second team took on a more foundational problem. CDS MS students Deepanshu Mody, Adhiraj Singh, Ghina Al Shdaifat, and Hamza Alshamy, working with Chris Tanner, Craig Schmidt, Adam Wiemerslage, and Varshini Reddy of Kensho, an AI arm of S&P Global, developed a new tokenization algorithm for greedy inference. Tokenization — the step that breaks text into pieces a language model can read — is one of the lowest layers of the modern LLM stack, and changes there ripple through everything built on top.
A third Fall 2025 team worked on a different UNDP project exploring how agentic AI can support procurement data preparation and audit review. CDS MS students Hejun Chen, Carina Sun, Kevin Chong Zheng, and Jiaming Lin, mentored by UNDP’s Mohamed A. Marrakchi, built a prototype that combines cleaning and feature-engineering agents with a RAG-based audit assistant. The workflow prepares messy procurement records, retrieves relevant evidence and policy context, and turns each purchase order into a structured, traceable review case for human reviewers. Its goal is not to replace audit judgment, but to reduce the manual burden of reconstructing evidence and make potential issues easier to inspect. The team has written up their work in two blog posts: Preparing Messy Data for Agentic Analysis: A Two-Agent Approach, and Building a Traceable Procurement Audit Agent: From Scattered Evidence to PO-Level Review.
“What stood out to me was how fast the team went from learning the tools to thinking like practitioners,” Marrakchi said. “For our office, working with students of this caliber is a preview of how auditors and engineers will work together, and they pushed our thinking as much as we pushed theirs.”
Capstone advisor and former CDS Moore-Sloan Data Science Fellow Anastasios Noulas has described the course as “a vital experience for students ahead of their graduation and prior to entering the job market,” and one that builds problem-solving and teamwork skills harder to develop in standard coursework. For partners, the benefits run in the other direction. Students bring recent technical training in current ML and NLP methods, applied statistics, and modern software practice to a problem the partner organization is already working on. That gives companies, nonprofits, and labs fresh perspective on real internal problems, early exposure to the CDS talent pipeline, and a working prototype or piece of research at the end of the semester.
The three Fall 2025 projects above show how broadly Capstone collaborations can range: applied computer vision for sustainable reef tourism, foundational NLP research, and an audit-support workflow shaping how a UNDP team thinks about AI-assisted review.
Submissions for the Fall 2026 Capstone Project are open through August 2, 2026. Companies, nonprofits, and research labs interested in proposing a project can submit through the Capstone Fall 2026 Project Submission Form or email ds-capstone@nyu.edu.
By Stephen Thomas
