Every hard money loan we make sits on the same seven risks: can the borrower pay, is the collateral valued correctly, is the property in usable condition, is the title clean, are the wires safe, is the borrower honest, are we compliant. Done well — really well — that's three hours of document review per file. Across a year of loans, that's a part-time job spent reviewing PDFs.
Most lenders our size cut corners on this. We wanted to do it more rigorously, not less. The question was how.
This month we shipped the answer: an AI underwriting system, built with Claude, that ingests a borrower's complete file plus any call transcripts, runs them against our seven-risk framework, and produces a structured pre-underwriting brief in ten to fifteen minutes. This used to take multiple hours and run the risk of human error playing a factor. This is an absolute game changer.
This article walks through what it does, the framework it follows, and what it produces. There's a downloadable sample analysis at the bottom — built on a fictional borrower, but identical in structure to what we run on every real file. The full skill text is embedded too, so you can read it inline or copy it.
The seven risks
Before we describe the system, the framework. These are the seven questions every loan answers — and the single thing each control exists to do.
| # | Risk | The one purpose |
|---|---|---|
| 1 | Borrower Can't Pay | Confirm current liquidity and a realistic exit plan |
| 2 | Collateral Valuation | Recover our basis in foreclosure at any lifecycle point |
| 3 | Property Damage | Insurance pays HCF's basis if the property is destroyed |
| 4 | Title & Lien Integrity | Valid first lien at closing, maintained through the loan |
| 5 | Wire Fraud | The wire goes to a verified escrow account, not a fraudster |
| 6 | Borrower Fraud | The borrower is competent and honest — selection IS the control |
| 7 | Regulatory Compliance | The loan is clearly business-purpose; the exemption is bulletproof |
Every loan brief maps to these seven. Every section of the analysis output corresponds to one of them. There is no ad-hoc underwriting at HCF — this is the framework.
The framework was deliberately stripped to one purpose per risk.
What we built
The system is a skill that runs inside Claude Code, our developer environment. The borrower's documents come from our loan portal; our system pulls them automatically once a borrower submits a file. The skill ingests every document, classifies it, and runs each one against the relevant section of the seven-risk framework.
Specifically, it does this:
- Pulls every document from the loan portal — Articles of Organization, Operating Agreement, Certificate of Good Standing, W-9, Personal Asset Statement, government ID, property photos, inspection reports, soft credit report.
- Reads each document in full. Not summarizes, not skims. Long-context reading on the actual content of every PDF.
- Runs an independent comp pull on the property — same county, recent sales, similar bed-bath-square-footage. Reconciles against the borrower's stated ARV and uses the lower of the two for all loan math.
- Pulls the Propstream comp printout — manually dropped in by me, since Propstream has no API. The skill checks the borrower's stated ARV against that printout, which doubles as a quality-control step on the data going into every brief.
- Runs a stress test at ARV minus ten percent — checks whether the deal still works if the resale market softens.
- Runs a public-records sweep — court records (state and federal), tax liens, judgments, bankruptcy filings, foreclosure history, business entity standing, professional licensing.
- Reads every call transcript — and surfaces the qualitative signal documents alone can't carry. We'll come back to this.
- Maps everything against the seven-risk framework — generating a status (covered / partial / gap / red line) for each risk with supporting notes.
- Recommends an LTP tier — 100%, 95%, 90%, or decline — and stress tests its own recommendation.
- Produces a branded PDF brief — six to eight pages, structured, internal use, ready for me to review and sign off.
The red lines
The framework also defines what we won't lend on, regardless of how strong the rest of the file looks. The skill is hardcoded to flag any of these as a decline before reaching the recommendation:
- Owner-occupancy intent (any signal)
- Refusal to sign the Business Purpose Certification
- Material dishonesty detected in the first call
- Bankruptcy in the last 3 years, open judgments, pending material litigation
- Can't verify the source of the down payment
- Entity not in good standing with the state
- Active-duty military service (SCRA complications)
- Purchase price more than 10% above the AVM with no justification
- ARV that exceeds the defensible comp ceiling by a material margin
- No prior deal experience and no credible deal analysis or takeout commitment
- Visible structural, mold, or fire damage on the video walkthrough
- FEMA Zone A or V without a flood insurance commitment
- Title company unwilling to issue a 2021 ALTA loan policy with required endorsements
- Wire instructions that can't be callback-verified
- Borrower pressure to skip any control
A red line in any of these categories means decline. It doesn't matter if everything else in the file is pristine. We don't size around them. We don't extend exceptions for these things.
The Plaud call recordings (why the call matters)
The single biggest underwriting upgrade isn't in the documents. It's in the call.
Every borrower's first call with HCF is recorded with a Plaud device — a small voice recorder that transcribes locally. There's no cloud roundtrip and no third-party processor in the chain. The transcript flows directly into the same analysis as the documents.
Why this matters: documents tell you the what — the entity name, the ownership structure, the prior LLCs, the assets. The call tells you the who. How the borrower talks about prior deals. Where they hedge. What they volunteer about their contractor. How they describe their exit. Which questions they ask back.
Documents can be doctored. Calls are pretty hard to fake while you're talking. And the qualitative signal — confidence, specificity, willingness to say “I made a mistake here” — is often the signal that matters most.
Most lenders rely on doc review alone. We rely on the call too. Both flow into the same brief.
Qualitative data has historically been hard to use systematically to make decisions, but with the AI models available today, it is.
Why Claude, specifically Opus 4.7
We use Anthropic's Claude — specifically the Opus 4.7 model — as the engine inside the system.
Opus 4.7 is the most capable model Anthropic has released as of this writing. The two capabilities that made this build actually work were long-context reasoning (the model holds the entire document package in one context window without losing track of cross-references) and structured-output discipline (the model produces the same brief format every time, with the same sections in the same order, regardless of what's in the file).
This wasn't possible a year ago. The previous generation of models couldn't read this much in one pass without dropping detail, couldn't hold a complex framework like our seven risks consistently across a long output, couldn't reliably distinguish what to flag from what was minor. Opus 4.7 can.
The hard rules
A few rules in the skill aren't judgment calls — they're hard limits the system enforces on every deal:
- 70% LTV cap. Loan facility (purchase loan + reno holdback) divided by ARV must be at or below 70%. The system declines anything above and won't size around it.
- Lower-of-two ARV. Whenever the borrower's ARV and HCF's independent ARV differ, the lower number is used for all loan math. Always.
- Gap rationale at 10%+. If the borrower's ARV and HCF's diverge by more than 10%, the brief must contain a written rationale for why HCF's number is more defensible. No silent overrides.
- Stress test at ARV − 10%. Every loan is run through a haircut scenario. Anything that breaks under stress gets flagged in the recommendation.
- LTP tiers, not negotiated. 100% (rare, needs an exceptional file), 95% (default for repeat borrowers with verified liquidity), 90% (newer borrowers or thinner files), or decline.
What every brief contains
The system produces the same brief format on every loan, with the same sections in the same order. That structural consistency is half the value — every file gets the same depth, no matter what.
The full output runs 6 to 8 pages and contains, in order:
- TL;DR — 2 to 3 sentences with the recommendation in bold; any red line hit is surfaced first
- Borrower Snapshot — entity, credit, liquidity, track record, lead source
- Documents on File — table of every document received and its status
- Public Records & Online Presence — courts, civil/financial, entity standing, reputation, license verifications
- 7-Risk Framework Mapping — covered / partial / gap / red line for each of the seven risks
- Recommended LTP tier — 100%, 95%, 90%, or decline
- Borrower's Stated Position — the deal as the borrower describes it
- Borrower's Comps — their picks
- Narrative & Character Signals — pulled from the call transcript
- HCF Independent View — independent comps and ARV pulled by HCF
- Reconciliation — borrower ARV vs. HCF ARV, gap rationale if material, lower number used for all loan math
- Loan Math + Stress Test — base case at the selected ARV, stress case at ARV minus 10%
- Risk Flags — over-leverage, thin margin, market issues, property issues, borrower-deal mismatch
- Recommendation — approve / approve with conditions / decline / needs more info
- Conditions — any deal-specific requirements
- Missing / Needs Follow-up — checklist of items still pending
- Sources Pulled — every document, query, and external source used
The actual skill
Below is the full skill text — the prompt the system runs against — with company-specific infrastructure redacted (loan-portal API endpoints, API keys, internal file paths, individual names). The framework, methodology, and decision logic are unchanged from the production skill. You can read it inline or copy it to study it offline.
A real example
The example below is a real sample brief. The borrower, the entity, the addresses, the comp data, and the identifiers are fictional — we invented them so we could publish the brief publicly without disclosing real borrower information. But the structure of this output, the framework it follows, and the methodology behind it are exactly what HCF runs on every loan we underwrite.
The example is a clean approve at 95% LTP — a flipper with three closed deals, $128,000 verified liquid, and a Northside Richmond purchase that comps out within six percent of the borrower's number. Every section of the brief maps to one of the seven risks. The fraud sweep cleared. The stress test flagged one yellow item (LTV breaches the 70% cap on a 10% ARV haircut) but the borrower's balance sheet provides carry capacity. The recommendation walks through that reasoning before recommending forward.
Download the full sample analysis (PDF)
The human still decides
The system produces a recommendation. The system does not make the decision. Every brief gets read end-to-end by me before any term sheet goes out — the recommendation is one data point, not the verdict.
What the skill actually buys us is speed at the structured layer — pulling documents, reading them, running comps, mapping to the framework. That work used to absorb hours per file. Now it takes minutes. The hours are still spent. They're spent where they belong: on the judgment calls and the relationship.
The throughput math
Three hours to twelve minutes is more than an order of magnitude. What does that actually unlock?
It means we can underwrite every file with the full discipline. There's no incentive to cut corners on a smaller deal because the analysis is fast either way. There's no excuse to rush a complicated deal because the system carries the load on the structured part. The human attention — mine — gets spent on the judgment calls, the relationship, the parts that matter most.
It also means we can scale the loan book without scaling the operations team. If we double our origination volume next year, the underwriting cost per file doesn't double. The lean operation isn't a constraint we tolerate. It's a design choice that gets compounded by every file we process.
What it means for you, if you invest with us
Three things.
First, every loan in our book is underwritten with the same framework, the same depth, and the same thoroughness. There's no diligence-quality drift as the book grows. Claude doesn't get tired on a Friday afternoon — the fiftieth file of the month gets the same rigor as the first.
Second, the qualitative signal is captured. Most private credit funds underwrite from a credit memo and a balance sheet. We underwrite from those plus a recorded conversation, a public-records sweep, and a structured framework that flags what's missing. More signal, fewer surprises.
Third, when something doesn't pencil, we say no — and we say it faster. Speed plus discipline isn't a tradeoff anymore. It's the design.