
Operational project · client anonymized
Site intelligence
Problem: A training business needed a disciplined way to compare where to open, staff, and grow instead of trusting one headline score.
Built: Decision support across 33,772 U.S. ZIP codes and nine data sources, with uncertainty bands, screening verdicts, maps, and an export for review.
Measured finding: Leak-free validation reduced the old model's apparent correlation from 0.81 to 0.21. The useful outcome was catching false confidence before expansion decisions used it.
Role: data product and full-stack developer · client identity withheld