3.0×how often a race-blind AI wrongly flagged bank-approved Black applicants vs. white
1.2×after our fix, keeping 97% of the model's prediction score (PR-AUC)
264,340real Georgia mortgage applications, 2022–2023
1.Pick where the home is
2.Compare the two AIs. Neither one ever sees race.
3.See this person at every address at once
Our fixed AI can't see the neighborhood at all, so here its answer can't change with the address by design. What matters is the
result across all 117,547 real 2023 applications: the bank-style AI wrongly flagged bank-approved Black applicants 3.0× as often as
white ones; our fixed AI brought that to 1.2× while keeping 97% of the PR-AUC score. The trade-off: white applicants the bank approved
are flagged a little more often (4.0% → 5.6%).
Both AIs run live in your browser: these are the exact trained XGBoost models (every decision tree), exported from Python and checked to give identical answers. Data: public HMDA mortgage records from the CFPB / FFIEC (ffiec.cfpb.gov). Built with pandas, scikit-learn, XGBoost and fairlearn; Claude Code helped write code and review methods. Code, data pipeline and write-up: github.com/kabirshah4/race-blind-isnt-fair.
Team KV Gods · Veritas AI High School Hackathon 2026 · Prototype, not financial advice.
Both AIs run live in your browser: these are the exact trained XGBoost models (every decision tree), exported from Python and checked to give identical answers. Data: public HMDA mortgage records from the CFPB / FFIEC (ffiec.cfpb.gov). Built with pandas, scikit-learn, XGBoost and fairlearn; Claude Code helped write code and review methods. Code, data pipeline and write-up: github.com/kabirshah4/race-blind-isnt-fair.
Team KV Gods · Veritas AI High School Hackathon 2026 · Prototype, not financial advice.