AI-Powered Loan Approvals at a Leading Bank

First Atlantic's unsecured-loan funnel was bleeding. Average time-to-decision was 12.3 days, application abandonment between submission and decision sat at 34%, and the bank had quietly tightened its policy cutoffs four times in two years to compensate for rising charge-offs in the 660-700 FICO band.

The existing underwriting stack was a brittle combination of a 2014-era logistic regression scorecard, a mainframe rules engine that only the original vendor could change, and 23 underwriters in Charlotte who manually reviewed every application above $35,000. Worse, the bank's last fair-lending review had flagged a statistically significant adverse-action disparity in two protected classes that nobody could fully explain.

The CEO's directive was sharp: get to same-day decisions on the bulk of the book, do it without raising loss rates, and produce a model risk dossier that the OCC examiners would consider exemplary -- not just adequate.

TekNinjas delivered an end-to-end MLOps platform on a hybrid GCP / on-prem footprint over an 11-month engagement. The work was scoped jointly with the bank's MRM (Model Risk Management) function from week one -- we wrote SR 11-7 documentation in parallel with code, not after.

Decisioning models

We trained two complementary LightGBM models -- a probability-of-default model and an early-delinquency model -- on 7.2 years of internal performance data enriched with Plaid cash-flow signals, FactorTrust thin-file attributes, and (for borrower opt-in) verified income from Argyle. Feature stores ran on Vertex AI Feature Store with a 90-day rolling backfill. Hyperparameter search used Vertex AI Vizier with strict separation between train, validation, and out-of-time test windows.

Fairness and explainability

Every model release passed a four-stage fairness gate: Disparate Impact Ratio under Equal Opportunity, Demographic Parity Difference, calibration drift across protected classes, and adversarial debiasing on the residual feature set. SHAP-based per-decision explanations were persisted to BigQuery and translated into Reg B-compliant adverse-action reason codes by a deterministic mapping layer that compliance signed off on.

Decisioning runtime

The serving layer ran on Cloud Run with a 95ms p99 inference budget. Real-time feature lookups were cached in Memorystore with a 30-second TTL. Borderline cases (probability-of-default in a calibrated grey zone) routed to a streamlined underwriter queue that surfaced the model rationale, the top five SHAP features, and a recommended counter-offer. Every override -- underwriter or model -- was captured and fed back into a weekly governance scorecard reviewed by MRM and Compliance.

The platform went live in production in March 2026 after a 90-day champion-challenger period running silently behind the legacy scorecard.

• 71% of applications now decision automatically in under 90 seconds
• 13.6% lift in approval rate at the same forecasted loss rate
• Application abandonment fell from 34% to 9%
• Underwriter capacity reallocated from L1 review to high-value exception handling and small-business loans
• Adverse-action disparity in the previously flagged segments closed below the 80% rule threshold
• OCC review concluded with zero MRA findings on the model risk dossier