For lenders and credit teams
Screening is mechanical. Credit judgment is not.
PSV builds the system that reads every borrower package on arrival, runs the screening pass against your credit box, and drafts the memo straight out of the documents. Every figure carries the page it came from. After close, it keeps watching the book.
Send one live package and your memo template. We will show you what comes back on your own file.
Where the week goes
The wide part is mechanical. The narrow part is the reason you employ credit people. Today both are being paid for at the same rate.
Built by institutional CRE teams and AI researchers previously at
You have been here before
Four things you have already tried on this exact problem.
None of them were stupid. Each one made a piece of the mechanical work faster, and none of them gave your credit judgment any more room.
- 01
You gave the credit team AI seats.
It summarizes a rent roll well enough. It has never seen your credit box, it does not know what this sponsor told you in 2023, and it cannot tell you which page the number came off.
- 02
You bought document extraction.
Fields come out. Judgment does not go in. Extraction is indifferent to whether the operating statement and the rent roll disagree, which is the exact moment a credit person needs to be interrupted.
- 03
You added analysts to absorb the volume.
Screening throughput went up and so did review. More packages get looked at, by people whose scarce skill is the credit call, spent instead on re-keying documents into a template.
- 04
You standardized the credit memo.
The template is consistent now. Filling it still means pulling the same figures out of the same PDFs, deal after deal, and re-deriving anything the last person could not show their work on.
Notice what all four have in common. They sit on top of the documents. Not one of them sits underneath the credit process, which is where the thing you actually need lives.
Why none of it stuck
The tools sat on top of the documents. Nothing sat underneath the credit process.
Three failures, and they are the same failure looked at from three angles.
Every package arrives a stranger
Nothing carries forward. Not your credit box, not the exception this committee approved last spring, not what this sponsor reported the last three times. Each file is underwritten as though it were the first one.
The output cannot be shown
A figure with no provenance is not an answer, it is a claim. Before it can go in front of a committee somebody has to open the PDF and find it, which is the work the tool was bought to remove.
The book goes quiet after close
Origination gets the tooling and the attention. Covenant testing, reporting, and watchlist work stay manual, so the portfolio view lags whatever is actually happening in the assets.
Your credit team is not slow. It is re-deriving.
A number nobody can trace is a number somebody has to rebuild. That is where the hours go, and it is why the scarce thing in the building, the credit call itself, keeps getting spent on screening and re-keying.
What was actually missing
One system underneath the whole book.
Not another tool beside the LOS. A layer beneath the credit process that already holds what your firm knows, so that every package that arrives is read by something that has met your last four hundred.
PSV calls that layer the Company Brain. On a lending desk it is the difference between an assistant that can summarize a rent roll and a system that knows this sponsor missed a reporting deadline in 2023 and that your committee does not do cash-out above a certain leverage without a sweep.
- What it holds
- Your credit box as you actually apply it, your covenant language, the exceptions your committee has approved and the ones it has refused, and every sponsor you have already lent to.
- What it does with it
- Reads each new package against all of that, rather than against the average deal. The screening pass, the extraction, and the first draft of the memo are mechanical work, so they get done at volume.
- What it hands you
- The files that deserve a credit person, with the reasoning attached and the sources still stapled to it. Judgment gets spent on the deals where judgment is the constraint.
Traceability
Every figure carries the page it came from.
This is the part that is hard to build and easy to promise, so here is the mechanism rather than the claim. Pick a line in the draft memo and the system shows you the document and the page it was read from. Where the package does not support a field, the field is flagged instead of filled.
Credit memo, draft
Illustrative
Source
TracedTrailing twelve month operating statement
Page 12
Net operating income and the debt constant are read off the trailing twelve, not off the pro forma the package leads with.
Field 1 of 5. In the deployed system this opens the document at the page.
A lender cannot act on a number it cannot trace. So nothing reaches a memo without its provenance attached, and the fields the package does not support come back flagged rather than quietly filled in.
Additive, not a replacement
We build on the accounts you already pay for.
Your loan origination system stays the system of record. Your seats stay your seats. We build the layer underneath them, and the work your team already does on those tools gets better because of it.
Microsoft Copilot
Already deployed across M365, already through IT and vendor review.
With the Brain
The credit work runs inside the tenant your auditors have already looked at, so nothing new has to be argued through procurement.
Claude
Where your analysts already read long documents and reason through them.
With the Brain
Answers arrive already holding your credit box, this sponsor's history with you, and the exceptions the committee has actually approved.
ChatGPT
Seats are live and two or three people have quietly gotten good at it.
With the Brain
The prompts that work stop living in one analyst's head and become something the credit team owns and can hand to the next hire.
If your firm has standardized on one of them, we build there. If credit runs one and the rest of the shop runs another, the Brain serves both, so your context does not fork with the tooling.
How this gets built
Three engagements. Most credit teams start with the first.
They are sequential more often than they are a choice. You cannot build the right thing before you know where the time is going, and you cannot train a team on a system that does not exist yet.
- 01
AI Audit
Where judgment is actually being spent.
We follow a package from the moment it lands to the moment somebody signs: what arrives, who touches it, which steps are mechanical, and where the credit call is being crowded out. The map is yours whether or not you build anything with us.
- 02
Company Brain and AI Deployment
The layer underneath, then the agents on top of it.
Your credit box, covenant language, exception history, and sponsor record encoded once, then intake, screening, memo drafting, and monitoring built on top and tested against files you have already decided.
- 03
AI Training
So the capability does not leave with one analyst.
PSV led training for the credit team on the system you deploy, plus the CRE AI Institute for the people on your desk who want to build their own tooling rather than wait for someone to build it for them.
The deployed system
What it actually does, from intake through the life of the loan.
Half of this list sits after close, which is the half most tools in this category skip. A credit book is not a pipeline that ends at funding.
01
Intake that never queues
Every borrower package is read on arrival, whatever shape it turns up in, and normalized into the same structure so two deals can be compared without a person retyping either one.
02
The screening pass
Each file is tested against your credit box and returned with the reason it landed where it did. The nos come back explained, which is what makes them safe to trust at volume.
03
A memo drafted from the documents
Your template, filled from the package rather than from memory, with every field carrying the document and page behind it and every gap flagged instead of quietly assumed.
04
Sponsor and guarantor history
What this borrower told you before, what they delivered, and where the current package departs from either. The institutional memory stops living in one credit officer's inbox.
05
The reporting calendar, enforced
Covenant tests, financial reporting, insurance, and tax deadlines tracked against what actually arrives, so a missing statement surfaces as an event rather than as a discovery.
06
Watchlist that watches
Post close performance read against the covenants as written, with the supporting page attached to every trigger, so the book is current rather than reconstructed each quarter.
What changes
We are not going to put a number on your desk that we cannot source.
This page argues that a figure without provenance is worthless, so it would be a strange place to publish one. No lender engagement of ours is publishable yet. Here is the class of outcome these builds are designed to produce, and the work you can already read in full.
Screening stops depending on who is free
Every package gets the same first pass on the day it arrives, against the same box, with the reason written down. The nos come back explained, which is the only way a no is safe to trust at volume.
The memo survives the committee
Questions about where a figure came from get answered in the room instead of in a follow up email, because the field is still attached to the document and the page.
The book is current, not quarterly
Covenant tests and reporting run against what actually arrived. Watchlist movement surfaces as an event with its evidence attached rather than as something discovered at the next review.
The engagements we can show, with the numbers their clients approved, are on the case studies page. Read those before you believe anything on this one.
Read the case studiesYour environment
Built around your environment, not ours.
Where the work runs is a deployment decision you make. It is not a constraint we impose, and it is usually the first question your risk team asks.
We will not tell you a piece of software carries your compliance obligation. What it carries is a legible record of what was read, what was produced, what was flagged, and who approved it.
- Runs in your Claude or Copilot workspace, or in your own cloud environment
- Agents scoped to the systems, folders, and deals you approve
- Every figure traced to the source document and the page it came from
- Unsupported fields flagged rather than filled
- Humans approve what becomes firm knowledge, and every step is logged
Before you take it to credit committee
The questions a credit team asks.
Straight answers, including the ones where the answer is no.
No. Your LOS stays the system of record and your credit policy stays where it lives. PSV builds the layer that reads the borrower package before the file is even opened, drafts the memo against your own template, and keeps watching the credit after close. Nothing about your book of record moves.
Every field in the draft carries the document it was read from and the page it sat on. Open the field, see the source. That is the point of the build, not a feature bolted onto it: a lender cannot act on a number it cannot trace, so nothing reaches a memo without its provenance attached.
The field is flagged, not filled. The system will tell you that the reliance letter is referenced and missing, that the rent roll and the operating statement disagree on occupancy, or that the guarantor schedule stops short of contingent liabilities. Silence is the failure mode we design against.
No, and it is not built to. It does the screening pass and the document work so credit judgment is spent on the deals that deserve it. The recommendation, the structure, the conditions, and the sign off stay with your people.
That is the normal starting condition for bridge and balance sheet lenders, and it is why the build is custom rather than a product. We encode your box, your covenant language, your exception history, and the questions your committee actually asks, then test it against files you have already decided so you can see where it agrees with you and where it does not.
Monitoring is the half most tools skip. The system tracks the reporting calendar, reads what the borrower actually sends in, tests it against the covenants as written, and raises the watchlist items with the supporting page attached. The book stays current instead of being reconstructed once a quarter.
The deployment runs in your own Claude or Copilot workspace, or in your own cloud environment. Borrower files, sponsor history, and credit precedent are never used to train public or shared models, and nothing crosses between clients. Per firm isolation is a hard guarantee.
Every read, every draft, every flag, and every human approval is logged and reviewable. We do not certify your program and we will not claim a regulator has blessed the software, but we do give your credit review, internal audit, and examiners a legible trail of what was read, what was produced, and who signed it.
In phases. First we study your intake, your memo template, and your credit box. Then the system runs in shadow mode against files you have already decided, so you are grading it on known answers. Only then does it touch live packages, and only in the seats you choose.
Lean teams feel the problem hardest, because there is no bench to absorb the flood and the same two people screen, underwrite, and monitor. The build scales down: start with intake and the screening pass, add the memo, add monitoring when the first two are earning their keep.
Let’s talk
Send us one package and your memo template.
We will run it and show you the draft, with every figure traced to the page it came from and every gap flagged. On your file, not a demo file.
We take on a limited number of new partners each quarter.
Prefer email? zed@pacificsoftwareventures.com
Guardrails built inAES-256 encryptedYour data stays yoursOngoing optimization included