Loan document AI for private lendersLoan files read, checked and summarized,
with borrower data kept under your control.
Your team stops re-typing numbers from bank statements, appraisals and title reports. AI pulls the figures, links each one to the page it came from, checks them against your lending rules, and sends anything uncertain to a person. The models run on infrastructure you control, so borrower documents are not sent to OpenAI or other AI model providers.
Loan servicing workflows for a private real-estate lender, with decimal-exact money handling and append-only financial history.
Titan Funding hired me to build loan servicing software. I built servicing and origination on one platform whose books reconcile to the penny. The client later chose to extend the origination tool it already used instead.
Read the private real-estate lending platform case study →- Scanned PDFs read with OCR and filed by building system, first build to production →
- A small model run locally through llama.cpp, no per-page API bill →
Lending software, document reading and local AI, each shipped for real clients. Loan document AI is a new offer that builds on that work.
For lenders whose loan files are read by hand and whose data cannot leave their control.
When lenders call me, it’s usually one of these.
From loan file to reviewed summary
Sort and extract
Documents come in by upload or email and are sorted by type. Balances, deposits, ARV, liens and terms go into fields, including scanned pages, each linked to its source page.
Run your rules
Your LTV, DSCR, reserve and document rules run on the extracted figures.
A person decides
Only what rules can’t decide goes to the model. Low-confidence fields and every flag go to a reviewer.
One-page file summary
For the underwriter, with every figure traced to its page.
Before: someone reads every page and types the numbers. After: they review the flags.
What gets built
Scoped to your loan products and documents, with safeguards and ownership included.
Document intake & sorting
Upload or email files in; each document is sorted by type.
Extraction with source-page links
Balances, deposits, ARV, liens and terms pulled into fields, including scanned pages, each linked to its page.
Your lending rules, run first
Your LTV, DSCR, reserve and document rules run on the extracted figures before any model is asked.
Review queue with confidence thresholds
Low-confidence fields and every flag go to a reviewer. Nothing is approved by AI.
Audit trail
Every extraction, edit and approval is recorded.
Hosting you control
Runs on your cloud account or server, sized in the Build Plan.
Hosted AI tool, off-the-shelf software, or private AI?
If a hosted tool or off-the-shelf software fits better, I’ll tell you on the call.
| Factor | Hosted AI tool or API | Off-the-shelf lending software | Private AI with The Forge |
|---|---|---|---|
| Where documents are processed | AI provider’s servers | Vendor’s servers | Models you host |
| Fits your lending rules | Generic | Partly | Built around your rules |
| Cost as volume grows | Per-request fees | Per-seat or per-loan fees | Your hosting cost, priced upfront |
| Traceable figures | Limited | Varies | Every field linked to its page |
| Best when | Low volume, non-sensitive data | Standard loan products | Sensitive data, custom rules, steady volume |
What the Build Plan gives you (from $500)
Loan Document AI: common questions
Yes. The models run on infrastructure you control, so documents are not sent to OpenAI or other AI model providers. Your cloud provider still hosts the server, the same as your other systems.
It depends on your documents, so it is measured on them during the build, not assumed. Rules handle routine checks, confidence thresholds send uncertain fields to a person, and every figure links to its source page.
There are no per-request fees to an AI provider. You pay for your own hosting, which the Build Plan sizes and prices upfront. It grows with volume.
Open models chosen for the documents and the hardware. In client work I have run Qwen 2.5 7B through Ollama and Phi-3-mini through llama.cpp.
No. It reads, checks and summarizes. People make every credit decision.
Yes. It can send data into your current system or a custom platform, including the kind of lending software described on the private lending software page.
You do. You own all code and IP on final payment, and the accounts and data stay yours.