This project will use the Vanderbilt Poll, a long-running initiative led by Josh Clinton, as the foundation for building an AI-powered platform that transforms how survey data can be explored and understood. By leveraging large language models, the tool will ingest questionnaires, toplines, and micro-data, then generate visualizations, suggest related questions from past surveys, and surface subgroup breakdowns—while enforcing statistical guardrails to ensure rigor. Starting with a decade of Vanderbilt Poll results, the project will establish a framework for future polling analytics, ultimately enabling any survey to be uploaded and interactively analyzed. The goal is to empower non-coders, researchers, and the public alike to discover insights, compare trends, and promote transparency in survey research.

Stuttering affects about 70 million people in world, approximately 1% of the world's population. About 5-10% of children stutter at one point in their lives, with a quarter maintaining their stutter throughout their lives. However, it has been shown that speech therapy in early stages can have upwards of an 80% success rate in alleviating stutter. In this project, the team will work with VUMC researchers to develop models that leverage audio, text, and other types of data to classify and better understand the nature of stuttering. This project aims to leverage automatic speech recognition models, audio models, and other multimodal models to enhance early diagnosis and inform treatment planning.

Multimessenger astronomy combines different signals from the Universe—such as gravitational waves, light, and neutrinos—to study extreme cosmic events. Gravitational waves, in particular, allow us to directly probe collisions of black holes and neutron stars, offering insights into the nature of gravity and matter under the most extreme conditions. However, gravitational wave signals are faint and buried in enormous volumes of detector noise, and connecting them with other messengers requires fast and reliable analysis. Artificial intelligence is becoming essential in this field: it can detect weak signals in real time, match them with observations from telescopes or particle detectors, and help reveal the full astrophysical picture. 

Project 1: This project will explore applying GW-Whisper, an adaptation of OpenAI’s Whisper model for gravitational wave data analysis. Students will help optimize and apply GW-Whisper to search for intermediate mass black hole mergers in gravitational wave data and use machine learning methods to estimate the physical properties of these cosmic events.

Project 2: In this project, students will build interactive AI-driven visualization tools to showcase how GW-Whisper and other machine learning models perform in detecting and analyzing gravitational wave and multimessenger signals.

Small language models can be trained to use reasoning to extend their capabilities significantly. We will explore the extent we can train models to perform novel reasoning tasks in specialized fields requiring challenging and unique reasoning. In this critical, foundational work, we will examine how far small language models can be trained to match, and perhaps exceed, much larger models on specific reasoning tasks.

This initiative advances two distinct but complementary projects: the next phase of the AI Legislation Tracker and the new AI Litigation Tracker, developed with Brookings’ Lawfare. While each tracker will stand independently, the litigation tracker will build on the methodologies, code base, and visualization frameworks established through the legislative tracker. Together, they will generate rigorous analysis and advanced visualizations of both legislative and judicial developments, creating mutually reinforcing tools that enhance understanding of the evolving AI legal and policy landscape.

Legislation Tracker - This project will focus on refining an existing code base - enhancing the existing tool with visualizations that help track legislative progress over time. Additionally, we will also build a deep research agent using LangChain, similar to the deep research functionality on ChatGPT, but focused on AI-related bills across the country. 

Litigation Tracker - This project will develop a new tool, leaning heavily on the existing Legislation Tracker codebase and work with various APIs and data sources provided by Lawfare. 

The National Security Pipeline Project is an automated evaluation framework designed to rapidly assess the capabilities, differences, and national security risks of new AI models and disruptive technologies. By combining technical capability analysis, AI-augmented brainstorming, and scenario-based simulations, the system not only identifies potential malicious uses—from cyberattacks and autonomous weapons to disinformation campaigns—but also characterizes how each new model or technology differs from prior iterations. High-risk findings are automatically routed to domain experts for deeper review. The vision is to establish the Vanderbilt Model Threat Assessment Center, a trusted, independent hub that provides authoritative evaluations of new models and technologies. With rapid-response briefs, in-depth analyses, and outputs formatted for both human and machine use, the pipeline delivers continuous, actionable intelligence to help national security leaders stay ahead of evolving threats.