From 3-day loan decisions to 8 minutes.
Replaced a 15-year-old manual underwriting process with an ML-powered decisioning engine — without touching the core ledger.
The Challenge
Manual underwriting was killing growth.
Meridian's underwriting team reviewed every loan application by hand. Three analysts, two days minimum, a dense spreadsheet model built in 2009. They were turning away 40% of applications not because applicants were bad risks — but because the analysts didn't have time to review them.
The CTO had tried to fix this twice before. Both times, a vendor sold them a black-box scoring tool that the compliance team refused to sign off on. They needed a model they could explain to regulators, auditors, and borrowers — in plain English.
We were brought in to build a decisioning engine they could actually own.
Our Approach
Build trust before building models.
The biggest challenge wasn't the ML — it was getting compliance, legal, and the underwriting team aligned on what 'explainability' meant in practice. We resolved that first.
Audit & Data Archaeology
We mapped five years of loan decisions, outcomes, and analyst notes. Found the 12 variables that actually predicted default. Discovered that 3 of the 30 variables in the old spreadsheet model were adding noise, not signal.
Explainability Framework
Before writing a line of model code, we designed the explanation layer. Every decision the model made would produce a human-readable summary: which factors were positive, which were negative, and what a borrower could do to improve their application.
Model Development & Validation
Gradient boosting model trained on 5 years of historical decisions, validated against 18 months of holdout data. Bias testing across demographics. The model outperformed the manual process on both accuracy and approval rate.
Integration Without Touching the Ledger
The core ledger system was untouchable — a constraint from legal. We built the decisioning engine as an independent service with a thin integration layer, so the underwriting team could use the new system without any change to downstream reporting.
Rollout & Human-in-the-Loop Design
Phased rollout: auto-approve obvious approvals, auto-decline obvious declines, route edge cases to human review. The underwriting team went from reviewing every application to reviewing the 15% that genuinely needed human judgment.
Results
What we delivered.
8 min
median loan decision time, down from 3 days
31%
increase in approval rate with no change in default rate
94%
of applications now handled without human review
"We'd been told AI underwriting wasn't possible in our regulatory environment. Anven Studios proved that wrong — and gave us a system we can actually explain to the OCC."
David Chen
CTO, Meridian Lending
Ready to Build?
Let's build your intelligent product.
Book a 30-minute strategy call. We'll listen to your idea, challenge your assumptions, and tell you exactly how we'd approach building it — no sales pitch, no commitment.