Situation

High-stakes domains like medicine need decision models that practitioners can inspect, trust, and explain, not black-box algorithms.

Task

Build an accessible R tool for constructing fast-and-frugal decision trees grounded in academic decision science.

Action

Designed and published the FFTrees R package with a companion algorithm, visualization tools, and an interactive web builder.

Result

In most datasets, FFTrees reduces the information needed to make a decision from hundreds of data points to 3 or fewer. Published in Judgment and Decision Making and cited over 100 times, with applications in radiation dose triage for lung cancer, lung cancer screening criteria, and clinical course prediction in respiratory disease.

Situation

A team of data scientists at a large pharmaceutical company relied on thousands of lines of fragile, undocumented SQL and SAS: slow to run, hard to audit, and inaccessible to non-expert users.

Task

Overhaul the team's end-to-end EHR data pipeline into a reproducible, human-readable system.

Action

Designed and led development of FlatironKitchen, an internal R package covering database connections, cohort derivation, variable creation, and visualization in a unified, human-readable API.

Result

Became the team's analytical backbone; presented at rstudio::conf 2020 to an audience of data scientists across industry. Standard analyses that used to take >100 lines of code now took less than 10. A senior epidemiologist reported: "It's like heaven."

Situation

A provider of longitudinal real-world claims data needed a way for clients to get fast, rigorous answers to common health economics and outcomes research (HEOR) questions, without waiting weeks for a bespoke study.

Task

Design and build an MVP interface for implementing custom cohort analyses following rigorous scientific standards.

Action

Led MVP design and development in collaboration with engineers, epidemiologists, biostatisticians, and senior leadership, building an R and SQL-powered dashboard with ICD/CPT-based cohort builders and configurable comparison arms.

Result

A fully functional prototype delivering clinical, demographic, cost, and outcome summaries on demand, giving clients a no-code path to decide whether a full study was worth pursuing. Average time to go/no-go decision dropped from 7 days to 30 minutes.