Now

Most of my time goes to client work at BLDS LLC. The firm has been doing quantitative discrimination analysis since the 1980s, starting in credit and insurance and expanding into employment, healthcare, and now AI more broadly. The work is usually downstream of a model that a client has built or is about to deploy. My job is to audit it, figure out where it disadvantages a protected group, and either propose a remediation or document the finding for whoever needs to see it. Sometimes that audience is the client's internal compliance team. Sometimes it is a regulator.

The part of the practice I'm most engaged with at the moment is the LLM and embedding-based work. A few of my current projects: a retrieval-augmented generation system built for a client; an embedding-based analysis of how an employment-screening model matches candidate resumes to job descriptions; and a small language model I'm training to detect discriminatory speech turn by turn, so each phrase can be flagged or denied at the moment of generation.

These problems are difficult in ways that fair-lending analysis on scoring classifiers is not. A scoring classifier produces a numeric prediction that you can stratify directly by group. A generative model produces text, and the question of whether one piece of text is more or less harmful for someone than another is harder than it sounds. There isn't a settled answer to what to measure across groups in the first place: refusal rates, sentiment of responses, hallucination patterns, performance on demographically marked prompts. Each captures something real and each can be gamed. A serious fairness analysis has to be honest about which one it is actually testing.

The rest of my time goes to more conventional fairness analyses on classifier models, internal methodology work at the firm, and reviewing analyses going out to regulators.

Last updated: May 2026.