Ihza MahendraResume
Back to Production ML at scale
Production ML at scale

Income-proxy model for credit-card eligibility

Scoring model that estimates user income from behavior to power credit-card approvals.

What it is

A model that estimates a stand-in for income from observed shopping behavior, used inside the credit-card eligibility decision.

What it's for

Tokopedia did not have direct income data on its users, but credit decisions still had to be made. A behavior-based surrogate gave the lending workflow a defensible signal to score against, with a low cost of false rejection.

How it was built

A LightGBM surrogate model trained on observed shopping behavior, calibrated against the lending team's accept/reject thresholds. Drove +0.75% true-positive uplift with minimal false negatives.

My role

Sole author.

Built with
LightGBMscikit-learnHive SQLAirflowGCP

Want the full technical depth, the tradeoffs, what broke, what I'd do differently? Ask the agent about this project.