Artificial Intelligence

Frontier AI Model Evaluation

I design and implement evaluation frameworks for large language models (LLMs) and multimodal AI models, with a focus on adversarial task design, failure mode identification, and measurement validity. My work draws on a PhD-trained background in causal inference and statistical methodology to bring rigorous measurement science to AI evaluation. Through Green Finance AI, I provide frontier model evaluation services for external clients, including adversarial prompt design and domain-specific benchmark development in economics, finance, and quantitative reasoning.

AI Alignment Research

My research agenda bridges computational social science and AI alignment, with a focus on the epistemological foundations of propensity evaluation: how do we measure a model’s dispositional tendencies toward misaligned behavior when ground truth is inaccessible and the data-generating process may be adversarially structured? I bring methodological tools from applied econometrics and illicit finance research, where I have spent over a decade measuring latent constructs under analogous conditions of strategic concealment and missing data.

Generative AI for Analytical Methods

I designed and delivered a 2-day agile workshop for the United Nations Department of Economic and Social Affairs (UN-DESA), equipping analysts and policymakers with hands-on skills in Generative AI (GenAI) for research and policy work. The training covered:

• Strategic adoption of GenAI – Understanding LLM architectures, strengths, and limitations.
• Prompt engineering & AI-assisted workflows – Practical applications for research, report generation, and data analysis.
• Privacy, ethics, and bias – Navigating risks and best practices in AI adoption.
• Live demonstrations & hands-on training – Custom AI tools for document processing, internal reporting, and automating analytical workflows.
• Custom chatbot prototyping – A proof-of-concept session on building internal AI assistants.

This workshop was designed to demystify AI, drive adoption, and provide practical, actionable skills for UN professionals.

Climate Change AI Summer School 2024

Slides

Climate Change AI Summer School 2023

AI for Global Climate Cooperation

I served on the jury of AI4GCC, a competition where teams designed negotiation agreements for climate change using a climate-economic simulation with AI agents.

Below is a video of the keynote speech I delivered.

AI for Sustainable Agriculture

I was selected as a Lead Machine Learning Engineer as a part of a team of 20 engineers across the world who use AI for good. The challenge was to develop an AI-based system to remotely monitor nitrogen-flow in farms in real-time.

I led the development of Agronomist AI, a context-aware chatbot that can translate a corpus of scientific papers on nitrogen modeling into actionable insights for farmers on the ground.

Twin Revolutions of AI and the Green Economy

Investigative Statistics for Dirty Money

AAAI 2023 Fall Symposium

I was on the organizing committee of a symposium on Artificial Intelligence and Climate: The Role of AI in a Climate-Smart Sustainable Future at the leading Association for the Advancement of Artificial Intelligence.

Wicked Problems

I have a working paper outlining the advantages of machine learning to study complex policy problems.

Complex policy problems like climate change and illicit finance require a diverse methodological repertoire and an agnostic approach to selecting the appropriate analytical tool to accomplish discrete inferential tasks. Drawing from the disciplines of political science, economics, and statistical data science, the project presented here tackles three distinct problems on causal evaluation, measurement, and missing data. This paper introduces the broad analytical lens of the dissertation and offers a perspective on empirical research for the study of complex real-world policy problems. The article argues that climate change and illicit finance can be understood as ``wicked problems'', and that doing so reveals the epistemological limitations of the common inferential framework that underlies much of the policy-relevant research in applied social sciences. Instead, this paper makes the case that machine learning approaches are uniquely suited to the study of “wicked” problems which resist systematic a priori formulation. The inferential framework of machine learning does not require us to accept that there is a simple generative model for the problem that can be known to be true. Instead, machine learning can be deployed in conjunction with domain knowledge to generate policy-relevant insights without requiring strong assumptions on the data-generating process in nature.