// Research portfolio

Building AI that reasons in the real world.

Active research programs across language models, knowledge graphs, network science, and AI for social good — at the intersection of rigor and impact.

AI & Complex Systems, interactive simulation — An encounter between artificial intelligence and a real-world complex system.

Active research

Human trafficking network visualization
01

Studies in Global Human Trafficking

A nine-year program mapping and understanding human trafficking networks worldwide. We develop AI-driven tools to extract entities, relationships, and locations from noisy web data, building interpretable, entity-centric knowledge graphs. Drawing on network science, NLP, and computational social science, we examine how trafficking organizations form ties, coordinate activities, and adapt across geographies and online platforms.

Selected work GlobalTrafficking.org ACM 2018 App. Net. Sci. 2020 arXiv:1712.03086 arXiv:1704.05569
· · ·
Neurosymbolic systems diagram
02

Neurosymbolic Systems for Real-World Domains

Bridging large language models with knowledge-based systems for flexible, reliable analytics in healthcare, public policy, and e-commerce. We explore how LLMs can be combined with structured knowledge graphs in a neurosymbolic framework — building domain-specific systems that users can query in natural language with no coding or formal logic required.

Selected work Expert Systems arXiv:2408.01935 arXiv:2409.10007
· · ·
DeepGraph network visualization
03

Network Modeling of Complex Phenomena

The broad availability of datasets in social media, corporate filings, and transportation has enabled exciting convergence of network theory and computational social science. This portfolio ranges from understanding policy impacts during COVID-19 to quantifying gender bias in literature and examining the structure of political campaign finance in the U.S. House of Representatives.

Selected work arXiv:2203.06491 arXiv:2307.07920 ASONAM 2023 Springer 2024
· · ·
Large language models illustration
04

LLMs & Cognitively Inspired Problem-Solving

We develop cognitively inspired benchmarks and methodologies to evaluate large language models on tasks involving decision-making, spatial navigation, estimative uncertainty, and commonsense behavior. Drawing from psychology and behavioral science, we focus on whether models exhibit human-like patterns in complex, real-world scenarios — moving beyond accuracy to assess how models act, not just what they predict.

Selected work Royal Society OS Springer 2023 arXiv:2407.01892 arXiv:2412.15501 arXiv:2405.15185
· · ·
Machine common sense diagram
05

Common Sense & Deep Cognition in Advanced AI

Toward reliable, knowledge-grounded reasoning in both LLMs and agent-based systems. We explore how advanced AI systems — large language models, RL agents, and hybrid neurosymbolic architectures — can be endowed with common sense and deep cognitive capabilities to reason effectively in open-ended, real-world environments such as medicine, law, and finance.

Selected work Nat Mach Intell 2022 Nature 2024 Sci Reports 2024 Nat Mach Intell 2024
· · ·
Knowledge graph in health
06

Knowledge Graphs in Public Health & Medicine

Rising health disparities are a defining social challenge. We build and apply knowledge graphs to deep questions in health and to power search systems like Kaiser Permanente's Finding Doctors and Locations engine. Current research extends to LLM-based chatbots that query these graphs naturally, plus work on network visualization.

Selected work SIGIR 2024 PubPub
Domain-specific knowledge graph
A1

Domain-Specific Knowledge Graphs

Since the Google Knowledge Graph was publicized in 2011, KGs have become central to neurosymbolic AI — from search to recommendations. In this interdisciplinary effort we built algorithms, frameworks, and tools for semi-automatic construction from raw data, applied across e-commerce, healthcare, and human trafficking.

Selected work arXiv:2307.12173 Eng App AI 2022 IEEE BigData 2015 arXiv:2310.05258
· · ·
Open-world learning game environment
A2

Open-World Learning

Funded by DARPA, we built OWL simulation environments based on games like Monopoly and Poker, and developed scientific approaches to better understanding structural violations of expectation in AI systems deployed in real-world environments.

Selected work IEEE TETCI 2022 Simul. Modell. 2021 arXiv:2103.00683