Applied ML · Credit risk · Capital One
I engineered and maintained the machine-learning suite behind auto-loan underwriting — models that decision loans at large scale by finding signal in the credit characteristics tied to loan performance.
This is a deliberately high-level account of proprietary work at Capital One — the problem, my role, and the approach, described qualitatively. It contains no confidential figures, internal systems, data, models, code, vendor or team names, or architecture.
The problem
Underwriting at scale needs models that separate genuine credit signal — the characteristics tied to how a loan actually performs — inside a heavily regulated decision process where the bar for validation and documentation is high.
What I did
I engineered and maintained the underwriting ML suite, and built and maintained the core loan-decisioning model at its center: feature and signal work grounded in loan performance, plus the ongoing model maintenance that a production decisioning system in a regulated setting requires.
Impact
The suite decisions loans at large scale, and the core model I owned drives substantial annual value. This is high-stakes applied ML where correctness, stability, and defensibility matter as much as raw accuracy.
What it demonstrates
Regulated, high-consequence applied ML: building and maintaining models whose decisions carry real financial and compliance weight, and measuring their value in dollars rather than offline metrics.