The AI-in-Gov Council’s Summer 2026 research program highlights a core priority of responsible public-sector AI: supporting focused, empirical work that can help government institutions better understand, evaluate, and apply AI in consequential settings.
This year’s featured projects reflect that mission across several domains. Together, they examine how AI can strengthen traffic-safety assessment, improve transparency into government AI use, stress-test high-stakes regulatory reasoning, and model complex transnational smuggling networks from difficult legal text. While the application areas differ, the projects share a common concern with evidence, rigor, interpretability, and real-world public impact.
Uncertainty-Aware VLLM-Assisted Multimodal Traffic Safety Assessment
PI: Anand N. Vidyashankar, Professor of Statistics, College of Engineering and Computing
Co-PI: Shanjiang Zhu, Associate Professor, Sid and Reva Dewberry Department of Civil, Environmental, and Infrastructure Engineering, College of Engineering and Computing
This project addresses an important public-sector challenge: how to move from traffic video and automated counts to safety assessments that are interpretable, auditable, and useful for transportation decision support.
Using VDOT traffic clips and a YOLO-grounded video analytics pipeline, the project combines structured visual evidence with vision-language model interpretation to analyze safety-relevant events such as crossings, yielding behavior, interactions, conflicts, and near misses. A central research contribution is its explicit treatment of uncertainty. Rather than presenting model outputs as fixed answers, the team is developing methods to measure instability, disagreement, and confidence intervals, and to surface review flags when evidence is ambiguous or unreliable.
The project is designed around an evidence-grounded workflow: detections, tracks, counts, and trajectory summaries form a reusable lower-level layer that supports higher-level safety interpretation. That structure helps make outputs more traceable and better suited for human review. The goal is not automated policy action, but decision support that helps transportation professionals screen locations, prioritize interventions, and evaluate risks with a clearer view of both model outputs and model limits.
Mapping What Matters in Government AI: Measuring and Visualizing AI Impacts
PI: Myeong Lee, Associate Professor, Information Science and Technology, College of Engineering and Computing
Co-PI: Sungsoo Ray Hong, Assistant Professor, Information Science and Technology, College of Engineering and Computing
As governments expand AI use, one of the most pressing questions is not only where AI is being used, but also how those uses can be measured, compared, and understood across agencies and policy domains.
This project builds an interactive visualization and analysis system for AI use in U.S. government agencies. The team is curating data from multiple sources—including agency inventories, government websites, code repositories, and news coverage—to build a more comprehensive and navigable view of the federal AI landscape. The work includes data collection, cross-checking and disambiguation, relevance review, and standardization so that fragmented public information can be turned into a more coherent research resource.
Beyond descriptive mapping, the project is also developing an AI impact framework that organizes use cases across dimensions such as economic effects, organizational value, productivity, justice and law, safety, and security. The result is not just a database, but a human-centered platform that can help researchers, policymakers, and practitioners explore patterns across agencies, application domains, AI types, and vendor relationships. By improving both visibility and interpretability, the project supports more informed discussion about how government AI is developing and what kinds of impacts it may produce.
Evaluating LLM Failure Modes in Disaster Recovery Regulatory Reasoning
PI: Catalina Gonzalez Dueñas, assistant professor, Sid and Reva Dewberry Department of Civil, Environmental, and Infrastructure Engineering, College of Engineering and Computing
Public-sector AI systems that support legal, regulatory, or benefits-related reasoning require more than general benchmark performance. They must be evaluated against the specific ways failure can occur in the domains where they may be used.
This project develops an open, reusable risk-auditing toolkit for large language models operating in disaster recovery and regulatory reasoning contexts. The team focuses on highly conditional and high-consequence policy text, including disaster recovery regulations and related eligibility standards, where seemingly small interpretive errors can create major downstream consequences.
The work centers on three major goals. First, it defines a domain-specific taxonomy of failure modes for LLMs processing disaster recovery regulations. Second, it constructs benchmark datasets that connect regulatory reasoning to real-world hurricane damage records and expert annotation. Third, it compares model behavior across adversarial stress tests and threshold-sensitive scenarios in order to better characterize robustness, calibration, and failure.
A particularly important contribution is the project’s multimodal benchmark design. By pairing regulatory text with annotated post-hurricane imagery and structured metadata, the team is building an evaluation framework that tests whether model outputs remain reliable when legal interpretation depends on physical conditions in the field. This makes the work highly relevant to public-sector AI assurance, where the challenge is not merely producing plausible answers, but documenting where and why systems fail before deployment.
Modeling Adaptive Migrant Smuggling Networks across Legal Regimes with Machine Learning
PI: Carlotta Domeniconi, Professor of Computer Science, College of Engineering and Computing
Co-PI: Guadalupe Correa-Cabrera, Professor, Schar School of Policy and Government
This project applies machine learning, knowledge graph construction, and graph mining to the analysis of adaptive human smuggling networks described in public case law.
The central challenge is that smuggling operations are large, heterogeneous, and highly dynamic, while the source material is noisy, unstructured, and narratively complex. To address this, the project automates the building and analysis of criminal-network representations from text. The team uses large language models and a knowledge-graph extraction pipeline to identify entities, relationships, and summaries from legal cases, then consolidates those results into structured graph representations for downstream analysis.
Those graphs support several forms of network analysis, including community and role detection, dynamic group modeling, event analysis, and anomaly detection. The broader goal is to better understand how smuggling actors and relationships evolve across time and across changing legal and enforcement environments. The project also emphasizes transparency and explainability by linking extracted relationships and findings back to source material.
A concrete case focus is the Laredo sector model, which the team uses to study patterns such as cross-border movement, stash-house dynamics, corridor usage, trailer-based transport, and the socioeconomic profiles of smuggled populations. In doing so, the project demonstrates how AI methods can help researchers study complex transnational legal and social systems while maintaining a strong connection to evidence.
Why These Projects Matter
These featured projects show that AI-in-Gov research is not just about applying AI faster. It is about building methods, tools, and evaluation practices that make AI more accountable, more interpretable, and more useful in real public-sector settings.
Across transportation, government transparency, disaster recovery, and the analysis of illicit networks, the projects share several commitments: grounding model outputs in evidence, making uncertainty visible, supporting human review, and designing research that can inform practice rather than simply demonstrate technical novelty.
That orientation is especially important for AI in government. In public-sector environments, the standard is not whether a system can generate an answer, but whether it can support better decisions with appropriate rigor, traceability, and oversight. These Summer 2026 projects offer strong examples of that broader goal in practice.
We are excited to feature these projects as part of the AI-in-Gov research community.