Mortgage needs more than ChatGPT
Naren and I started Balerion with the thesis that the next category-defining company in mortgage technology would sit at the intersection of three distinct disciplines:
Deep domain expertise from mortgage veterans
World-class AI engineering from leading research institutions
A data-driven forward-deployed motion ensuring lender adoption and ROI
Six months later, we’re proud to share that we’ve assembled that exact team, and every day we spend partnering with our customers reinforces our conviction even further. Mortgage is a complex, regulated domain that touches real people and real data. Building in this space requires not only understanding the intricacies of the domain, but also leveraging the latest AI models to solve deeply technical problems in a way that is secure, scalable, and designed to fit seamlessly into how leading lenders operate today. This blog post aims to share some of our initial learnings with the outside world.
The Challenge
Homeownership is a defining pillar of the American Dream, and it starts with applying for a home loan. Yet few understand the complexity that goes on behind the scenes in the loan manufacturing process, including those that have taken out a loan or several. Let’s walk through some of the numbers.
Disparate Data Sources: At closing, the average loan document package is ~750 pages. Those pages come from a variety of sources–structured IRS transcripts, semi-structured bank statements, and unstructured CPA attestations–and in a variety of forms, including electronic, photocopied, and hand-written.
Long-Horizon Planning: Data churns over the industry average of 42 days from application intake to closing: new documents like recent credit pulls are added, dated documents like verifications of employment and bank statements expire, fees must be re-disclosed.
Guideline Reasoning: Simultaneously, that 750-page file must be cross-referenced against 1200 pages of Fannie Mae underwriting guidelines. Or 2800 pages for Freddie Mac. Or 600 pages of VA guidelines. Or all of the above. Now do that in real-time as a loan officer is on the phone with a borrower, or across thousands of loans in a correspondent QC workflow where every added day of dwell time is a monetary loss measured in bps.
Custom Fine-tuning: Keep in mind this is only what’s standardized. It does not include custom lender-specific overlays, Non-QM guidelines, in-house correspondent QC rules, or ever-changing government regulations.
This is the messy reality, and why teams of just AI experts or just mortgage veterans have struggled to get a foothold thus far. The former lack the deep domain knowledge to wade through the complexity and seek out what matters and the latter lack the AI expertise to capitalize on the resulting findings.
The Opportunity
Grounding us once again where we began, the white picket fence and homeownership is close to every American’s heart. In practice, that means 2T+ in loan volume originated annually. Over 15,000 loans closed every day. An average cost-to-originate of $12,000 that is quickly creeping upwards. A widening gap between lenders stuck in the past of manual processes and lenders that have successfully deployed AI automation in their businesses. The challenge is real, but the opportunity is massive.
Our Approach
We started by building a foundational team. We assembled a world-class engineering team of UC Berkeley AI Engineers, Data Scientists, and Infrastructure Specialists. Then we paired them with industry veterans like Patrick Harkins who has been selling mortgage technology for decades starting at Ellie Mae and Devin Daly, former CRO at TRUE and Managing Director at Homebridge Financial (now CMG Financial) - a top 5 IMB in the country.
Then we built the technology: a proprietary full-stack UW system for Fannie/VA/NQM that is easy to customize and handles guideline ambiguity – powered by an LLM compiler that turns processes living in a pile of institutional documents into clean rails that AI automation can run on. We then secured this automation with guardrails to protect against hallucinations and packaged it in fault-tolerant infrastructure that can scale to thousands of loans and millions of pages in flight at once.
Finally we crafted the delivery motion. What good is the latest and greatest technology if it can’t meet lenders where they are at today? If it can’t demonstrate ROI? To this end, we assembled a data-driven forward-deployed team to embed in the nation’s top lenders and show what our technology can do on their own data. After all, seeing is believing. We paired data scientists with forward deployed engineers to sift through data and operational processes and build customized evals that point out where our agentic automation can deliver value on Day 1.
Looking Ahead
What this concretely means for a lender is fewer touches per loan file, and more confidence earlier in the process. It means empowering your team with AI to cross-reference, condition, stip, and clear even before a file hits UW in retail or as a QC check when it’s bought in a correspondent channel.
Our goal at Balerion is to continue making this vision a reality and safeguard the American Dream of homeownership. We are already in production with industry pioneers such as David Brecher and FM Home Loans originating $2B+ in loan volume. If you are interested in joining our team or a lender looking to streamline your operational processes, reach out at info@balerion.ai.
