Studying one hard problem deeply, and building methods that travel

The problem

Thousands of people need a kidney, and a living donor is often their best chance. Becoming a donor is a long process, though, and many willing people drop out or are ruled out late, after months of tests. Much of what donors actually experience is shared in public, in forum threads and social media posts, rather than in clinical records. My research asked what those voices could teach transplant programs, and how the evaluation process itself could waste less of a donor's time.

How I approached it

I started by learning the field from the people in it. I worked alongside transplant clinicians, social scientists, and qualitative researchers, and I studied the published literature until I could describe the problem in their terms. Only then did I choose methods, and I chose them by researching what had already been shown to work.

Whenever a solution worked for one case, I asked what it would take to make it work for the whole class of cases. That habit produced most of the results below.

Methods that generalize

• A repeatable way to turn expert judgment into ground truth. I converted qualitative concepts from domain experts into labeling criteria and adjudicated labels for more than 200,000 texts. That process became the basis for classifiers with F1 scores above 85 percent on two separate projects, and it applies to any field where experts can recognize something that is hard to define.
• Promptimizer. I treated prompts for language models as combinations of features and searched them with genetic algorithms. The method beat Google DeepMind's PromptBreeder on the GSM8K math benchmark while running a smaller open-source model on a single machine.
• Dialogue Until Classification Consensus. Two AI agents receive different instructions, argue their positions, and revise until they agree on a label. It improved on chain-of-thought prompting by up to 10 percent and earned an honorable mention at the INFORMS Healthcare Conference.
• A new optimization problem for medical testing. I defined the Single-Dependence Sequential Testing Problem, which orders a donor's tests so that disqualifying results surface early. Simulation showed when a simple rule performs nearly as well as exact optimization, which is a finding any team that sequences inspections or checks can use.
• Testing a published conclusion. I replicated a social network analysis of an illicit organ trading network, re-engineered it, and published findings that challenged the original interpretation.

What this means for a client

Most problems a business brings me are one visible instance of a larger pattern. I bring the same discipline from my research: learn the field from the people in it, find out what the literature already knows, and build a solution that fixes the pattern so the next instance never needs a special fix.

Publications

• "The Single-Dependence Sequential Testing Problem for the Living Kidney Donor Workup Process." Soft Computing, 2025.
• "Automated Classification of Living Kidney Donation Experiences on Reddit." JMIR AI, 2025.
• "Social Network Analysis of Illicit Organ Trading Networks: The Medicus Case." Trends in Organized Crime, 2023.
• "Classifying Comments on Social Media Related to Living Kidney Donation." JMIR Medical Informatics, 2022.