The challenge
In non-traditional lending, every lender publishes its own guidelines: complex PDFs that can run 50 to 200 pages. A broker working with 10 to 15 lenders had to manually open and read PDFs for every client question, remember which lender allows what, call lender representatives and wait for clarifications, and compare 5 to 10 lenders side by side by hand.
A single qualification question could take 30 to 90 minutes of research across multiple documents. It was slow, error-prone and did not scale.
What we built
An AI-powered lending intelligence platform. Lender guideline PDFs are uploaded once, and from then on anyone on the team can ask plain-English questions and get instant, accurate answers, like having a lending expert who has memorized every guideline book and is available 24/7.
The platform holds every lender's guidelines as a searchable knowledge base. When a user asks a question, it understands the intent, breaks it into sub-questions, searches for the most relevant rules from the right lenders, and filters lenders automatically based on the loan's details.
A layer of specialized AI agents then produces the answer: fast single-lender chat for quick lookups, a deep-dive researcher for one lender's full program lineup, and a marketplace researcher that compares every eligible lender in parallel and ranks the options in seconds, work that would take a human analyst one to two hours.
How it works, in plain English
A lender's guideline PDF is read, its rules extracted from text, tables and matrices, and converted into a searchable knowledge base.
"Can I use rental income for a self-employed borrower with a 680 FICO?" is understood, broken into sub-questions, and matched to the right lenders.
Lenders that cannot possibly approve the loan, on credit score, loan-to-value or property type, are excluded before research even starts.
Eligible lenders are researched simultaneously, so comparing ten lenders takes the same time as looking up one.
Results
| Metric | Before | After |
|---|---|---|
| Answer one guideline question | 30 to 90 min | Under 30 seconds |
| Compare 5 lenders for one borrower | 2 to 4 hours | Under 2 minutes |
| Onboard a new lender's guidelines | 1 to 3 days | 15 to 30 minutes |
| Staff dedicated to guideline research | 1 to 2 people | Zero |
- 200+ lender guideline PDFs indexed
- 83%+ benchmark accuracy on real qualification questions
- Multi-tenant, each client's data fully isolated
Technical highlights
- -Hybrid search combining keyword and semantic, meaning-based matching for highest accuracy.
- -Three specialized AI agents, each tuned for a specific research task.
- -A modern vector-search and frontier-LLM stack, built multi-tenant from day one.
Why it matters
Senior staff who once spent 20 to 30% of their day on guideline research now spend under 5%, and the same team can support several times more loan officers without adding headcount.
The platform does not replace loan officers, it removes the tedious research work so they can focus on client relationships and closing loans. Client identity kept confidential.
Client identity kept confidential.