Skip to main content
Financial services
Financial services · Lending

ML decisioning with fair-lending evidence on day one.

We build and operate AI decisioning, document intelligence, and origination workflows for banks and non-bank lenders, with the ECOA / Reg B and model-risk documentation your examiner expects delivered with the system.

Lending outcomes

62%
Application throughput lift

Same underwriter team, post-decisioning rollout

<2%
Adverse-action disparity

Across protected classes, quarterly review

4.1x
Document intelligence lift

Pages reviewed per hour, vs. manual

90 days
Pilot to production

Median, scoped origination workflow

Engagement model

How we work with lending teams.

A four-phase path from portfolio review to a scoped decisioning slice in production with an examiner package in hand. The same principals across all four phases.

012 weeks

Portfolio and fair-lending review

Loan book inventory, current decisioning logic, and a written gap analysis against ECOA / Reg B disparate-impact expectations and your prudential examiner's most recent guidance.

024 weeks

Pilot scope and model risk framework

Narrow first decisioning slice (often prime, unsecured), board-ready model risk management memo, override workflow, and the adverse-action notice template aligned to the new model output.

038 weeks

Production build with examiner artifacts

Build and ship inside your tenant, with model cards, disparate-impact testing reports, change logs, and the third-party vendor risk package delivered alongside the system.

04Continuous

Operations and periodic validation

Monitoring, drift detection, quarterly fair-lending review, and annual model validation memos signed by a PRR principal. We own the run-rate; you own the credit policy.

Regulator alignment

Named frameworks, named deliverables.

Every framework below maps to a specific artifact we hand you, with a written owner and a refresh cadence. No checkbox theater.

ECOA / Regulation B

Equal Credit Opportunity Act

Disparate-impact testing reports, adverse-action notice templates, protected-class variance review, and the written validation memo for every deployed scorecard.

NCUA lending guidance

Member-business and consumer lending

Concentration risk documentation, member-business lending policy alignment, and the override-rate analytics your supervisory committee expects.

FFIEC Model Risk Management

Development, implementation, use

Model inventory, periodic validation, challenger model documentation, and the change-management trail from version one through production.

NIST AI RMF 1.0

Govern · Map · Measure · Manage

Govern, Map, Measure, and Manage function coverage with a profile document per deployed decisioning model.

From the field

One we shipped this year.

Regional bank · consumer unsecured book

Replaced a generic vendor scorecard with a governed ML decisioning system inside the bank's AWS tenant — quarterly fair-lending review and adverse-action notice path approved by compliance before the production cutover.

Talk to the lending team
62%

Throughput lift, same team

Pilot quarter, prime book only

FAQ

Questions your lending CTO is going to ask.

Honest answers to the five we hear most often on the first technical call.

How often do you re-validate the decisioning models you deploy?
At deployment, then quarterly. Every validation cycle produces a written memo with population stability, performance against challenger, and disparate-impact analysis across protected classes. The memo is signed by a PRR principal and lands in your shared evidence folder.
How do you generate adverse-action notices when an AI model declines an application?
Every deployed scorecard ships with a reason-code mapping that translates model output into ECOA-compliant adverse-action reasons. The notice template is reviewed by your compliance officer before deployment; we do not ship a model into production until the notice path is signed off.
How do you test for disparate impact across protected classes?
We run pre-deployment and quarterly post-deployment tests using your loan book by race, ethnicity, sex, age, and marital status where the data permits. Variance is reported with confidence intervals; anything over your written tolerance triggers a model freeze and review with credit policy and compliance.
What does the override workflow look like for underwriter and reviewer roles?
Every model decision carries a confidence band and a reason-code stack. Below your threshold, the application routes to an underwriter with full feature context and the contributing factors. Overrides are logged with reviewer ID, justification, and a post-decision audit queue your QC team samples weekly.
What artifacts do we hand the prudential examiner on day one of the next exam?
A binder, not a scramble. Architecture-of-record, model inventory, validation memos, disparate-impact reports, vendor risk packages for every third-party AI component, change logs, override logs, and the board-approved model risk policy. We co-author the response to examiner questions on request.

Why PRR

Why lenders choose PRR over the alternatives.

Fair lending is a deliverable, not an audit.

Disparate-impact testing, reason-code mapping, and adverse-action notice templates ship with the system on day one — written by the engineers who built the model, not a compliance team backfilling six months later.

We ship the workflow, not just the scorecard.

A model is the easy part. The hard part is the override queue, the post-decision audit sample, the QC dashboard, and the data feed that keeps the model honest in production. We deliver all of it.

Senior architects build what they scope.

The principal who writes the SOW writes the first commit. There is no handoff to junior staff after the contract signs, because there is no junior staff layer to hand off to.

Bring us your hardest lending decisioning question.

Thirty minutes with a principal. We will walk through your portfolio, the regulator on your back, and what a 90-day pilot would actually look like.