James Michelson

I am a research scientist with over a decade of R&D experience across industry and academia. My scientific expertise spans insurance and market design, polling and political analytics, AI safety, causal inference, and the philosophy of science—often with a focus on measurement and incentives. I am happiest when solving wicked problems and care most about helping organizations make high-stakes decisions under uncertainty. My resume is here.

I currently work as a software engineer for CFC Underwriting. Prior to this, I completed my PhD in Computation, Logic and Methodology at Carnegie Mellon University in computational microeconomics and philosophy of science (see Writing section below). I also spent a number of years working as a Data Scientist in polling and political analytics at Civis Analytics, Priorities USA, and Vox Pop Labs.

Recently, I've been enjoying using AI and configuring agents to tackle some fun problems, including:

Applying numerical analysis techniques to the auction theory conjectures derived from the simulation chapter of my PhD (here)
Working through sensitivity analysis of unmeasured confounders in observational causal inference time-series data (here)

I also like to cook. My recipes are here.

Writing (Academic)

Reflexive Measurement (2023) — a systematic account of the causal relationship between measurements and the data they elicit in the social sciences, arguing that scientific knowledge of the social world requires an explicit model of how our instruments causally affect what they measure.

Optimal Multi-Dimensional Auctions: Conjectures and Simulations (2022) — with Alexey Kushnir. Uses simulation to test open conjectures on optimal multi-dimensional auction design, and introduces the first open-source Python library for multi-dimensional auction simulations (code here).

Developing a Philosophical Framework for Fair Machine Learning: Lessons From The Case of Algorithmic Collusion (2022) — argues that existing fairness metrics, built around classification and protected-class discrimination, cannot capture harms like algorithmic collusion that affect all consumers, and proposes a normatively-scoped framework for extending fairness research to these domains.

Writing (Popular)

How To Legally Play Poker (If You Must) In The United States (2022) — casino poker is illegal in 48 states, so this note sketches a loophole: put poker on the blockchain, make the whole program (including the random seed) public, and it stops being a game of chance and starts being a game of skill — namely, the skill of out-computing everyone else at reproducing the deal.