Director of Product (June 2021 - February 2023)
Carputty was building a next-generation auto line of credit platform designed to provide fairer, more transparent vehicle financing through technology and data-driven underwriting.
The Business Context
Auto lending in 2021 was ripe for disruption. Traditional lenders relied heavily on stated income and credit scores, which often disadvantaged non-W2 workers like contractors, freelancers, and small business owners. Carputty's vision was to use actual bank data to make more accurate, bias-free lending decisions.
The challenge: How do you automate income verification at scale while maintaining accuracy and regulatory compliance?
The Manual Process Problem
When I arrived, Carputty's underwriting team was manually analyzing bank statements in Excel spreadsheets - a process that was:
Each application required 30-45 minutes of manual review
Different underwriters categorized income differently
Manual data entry and calculation mistakes were common
Growth required proportional headcount increases
Human judgment introduced potential for unconscious bias
This manual bottleneck was preventing Carputty from achieving their vision of fast, fair, automated lending decisions.
Building the AI Solution
Phase 1: Understanding the Domain
I spent weeks working alongside underwriters to understand exactly how they analyzed bank statements:
- •What patterns indicated reliable income vs. one-time deposits
- •How they handled irregular contractor payments
- •What red flags suggested inflated or fraudulent income claims
- •How regulatory requirements shaped their decision-making
Phase 2: Designing the ML Architecture
Working with our engineering team, I architected a solution that:
ML models identified income types from transaction descriptions and amounts
Low-confidence predictions were flagged for manual review
Underwriter corrections fed back into model training
Every decision was logged for regulatory compliance
Phase 3: Building the User Experience
The interface had to work for underwriters who weren't technical:
AI showed its best guess for each transaction
Underwriters could easily override incorrect categorizations
Visual cues showed which predictions to trust vs. review
Clear income calculations with supporting evidence
Technical Implementation
Data Pipeline
Plaid integration for direct bank data access and real-time transaction analysis
Real-time transaction categorization using NLP and pattern matching
Confidence scoring for automated vs. manual review routing
ML Models
Transaction classification based on description text and amount patterns
Debt-to-Income calculation and income stability analysis for non-traditional income
Fraud detection for unusual or suspicious patterns
Regulatory Compliance
Complete audit trails for all automated decisions
Human review requirements for edge cases
Documentation supporting fair lending practices
The Results
The impact was transformational for Carputty's operations:
From 30-45 minutes per application to 2-3 minutes for AI-reviewed cases
High-confidence cases required no human intervention
Complete straight-through processing for clear-cut cases
Business Impact
Enabled rapid scaling without proportional headcount increases
Improved consistency and reduced human error in income calculations
Positioned Carputty as technology leader in automated underwriting
Created competitive advantage for loan resale in secondary markets
Customer Experience
Dramatically reduced time from application to decision
More objective, data-driven loan approvals
Better outcomes for non-traditional income borrowers
Press Recognition
"By harnessing the power of AI, we provide an enhanced user experience and make finance decisions based solely on objective data, free from any influence of subjectivity, misaligned incentives, or biases."
— Patrick Bayliss, CEO, Carputty
Read full press release (September 2023)Strategic Lessons
Domain expertise is crucial
The solution only worked because I deeply understood the manual process first. You can't automate what you don't understand.
Human-AI collaboration beats pure automation
Rather than replacing underwriters, we augmented their capabilities - they could focus on complex cases while AI handled routine decisions.
Regulatory considerations shape product design
In financial services, compliance isn't an afterthought - it has to be built into the core architecture from day one.
Measurable impact drives adoption
Having clear productivity metrics made it easy to demonstrate value and get organizational buy-in for the new process.