What can LLMs tell us about the ETI?
2025We investigate how large language models perceive and simulate behavioral responses to tax policy changes, with implications for using LLMs in policy analysis.
Projects
Software I've built and research I've published — mostly at the intersection of policy, economics, and AI: what the law says, what it does, and what happens next.
Founder & CEO · 2026—
The Axiom Foundation
Publishes the law as open, verified code — every value citing its statute, every clause carrying its effective date, every encoding checked against independent implementations. Starts with tax and benefit policy; launches publicly July 28, 2026, with PolicyEngine as its first and most demanding consumer.
Co-founder & CEO · 2021—
PolicyEngine
Free, open-source software that computes taxes and benefits for any household and simulates reforms across whole populations. Policymakers, researchers, and benefit navigators use it in the US, UK, and Canada. See the research portfolio at policyengine.org/us/research.
Founder · 2019–2023
UBI Center
Open-source think tank researching universal basic income and cash transfer policies. The team produced over 60 reports on UBI proposals, child allowances, and carbon dividends.
AI eval
PolicyBench
Open benchmark scoring AI models on tax and benefit law, graded against exact calculations.
MicrodataPopulace
Calibrated survey microdata for policy simulation — national to local, fully reproducible.
Dev toolsCodeStitch
Combines GitHub repos, PRs, and issues into single files for AI context.
Python
microdf
Pandas extensions for weighted survey microdata, used across PolicyEngine's models.
macOSOpenMessage
macOS Google Messages client that gives coding agents access to your texts.
VS CodeTerminalGrid
VS Code extension for keyboard-driven terminal grids, built for agentic coding.
macOSRAMBar
Native macOS menu bar app tracking memory across Claude Code, VS Code, Chrome, and Python.
ChromeScrollywood
Chrome extension recording smooth-scroll videos and GIFs of any webpage.
Pythonopencollective-py
Python client for the OpenCollective API, with CLI and MCP server.
Monte Carlo
MacKenzie Scott's giving, in QALYs
GiveWell-style Monte Carlo model estimating the health impact of $26.3B in giving, running entirely in your browser.
ExplorableA year of ChatGPT is a fifth of a beer
Prices drinks in AI queries, and lets you flip every assumption behind published AI water numbers.
CivicVoter guides
Recommendations for California elections — propositions, candidates, and local measures.
Enhancing survey microdata with administrative records
November 2024A novel approach to microsimulation dataset construction that combines Current Population Survey data with IRS administrative records for improved tax-benefit modeling.
AI model policy impact forecasts: A narrative prompting approach
2024Examining the use of large language models for forecasting policy outcomes under different U.S. presidential administrations using narrative prompting techniques.
The double distortion: Income-based traffic fines and labor supply
2025Agent-based simulation analyzing welfare effects of income-based vs flat traffic fines, showing labor supply distortions from implicit marginal tax rates.
The disutility of uncertainty
2025How uncertainty around marginal tax rates affects social welfare.
What Nut?
2025Monte Carlo analysis of life expectancy from nut consumption, with evidence-traced claims and hierarchical nutrient models.
Beer price controls at Yankee Stadium
2024An economic analysis of consumer welfare, revenue, and externalities from stadium beer pricing policies.
The Mind: LLM coordination
2025Simulating The Mind card game with large language models to study emergent coordination without explicit communication.
Smaller experiments — games, prototypes, one-offs — live in the lab.