Case study · LoanSnap · 2019 – 2020
A mortgage in thirty-six hours
LoanSnap wanted to decide a mortgage with a model instead of a queue. I owned the origination workflow and the borrower experience. The clock went from fifteen days to thirty-six.
- Role
- Product Lead, reporting into the founding team
- Dates
- July 2019 to March 2020
- Surface
- Origination workflow, borrower experience, underwriting automation, a mortgage-backed investment product
- Users
- Borrowers, and the mortgage loan officers and processors working their files
The clock was the product
A mortgage is a document pipeline wearing the costume of a product. Income, assets, employment, title, insurance, and a list of conditions that only appears once an underwriter has read the file. Fifteen days from application to a decision the borrower can act on was normal for the industry, and normal was the thing LoanSnap existed to argue with.
Fifteen days is not slow in the abstract. It is slow against the things that expire while it runs. A rate lock. An offer deadline in a competitive market. A borrower's willingness to keep answering emails about a bank statement they already sent. Every day the file sits is a day the loan can be lost to a lender who is not faster, only earlier.
The pitch was that a lender reading your whole financial picture can tell you whether a loan is actually good for you. That pitch does not survive a two week wait for the answer.
What changed
Application to a decision the borrower can act on
Same loan products, same regulatory requirements, same underwriting standard. The work was in the sequence, not in skipping steps.
The sequence, not the screens
The redesign started with the workflow and got to the interface second. Three decisions did most of the work.
- Collect for the decision, not for the file. A loan stalls when a document is requested after somebody notices it is missing. Anything the decision depends on gets asked for while the borrower is still in the flow and still motivated, not two days later by email.
- Raise conditions at submission, not at review. Conditions are the hidden clock in a mortgage. When the system can generate them the moment a file arrives, the borrower and the loan officer start clearing them immediately instead of waiting for a human to read the file and then start the waiting again.
- One status view both sides trust. When a borrower cannot see where the loan stands, chasing status becomes the workflow, and every call interrupts the people who could be moving the file.
Then the automated underwriting system went in on top of it. It halved the time to fulfill borrower conditions and halved processing time, which is where the last of the fifteen days had been hiding.
The exercise came before the job
The work at LoanSnap started as a take home. In May 2019 I was given the qualification problem, which is the one every digital lender has and few will discuss in public. A model says no. A quarter of the people it says no to walk into a branch and get approved by a human. Those borrowers are not bad risks. They are borrowers whose evidence does not fit the form.
I wrote it up as a product, not as an essay. Two personas, the borrower and the mortgage loan officer. Eight user stories ranked on user value against effort. The one I took first was the borrower with a good job, a real income and no credit history to prove either.
Then five ways to solve it, scored against each other rather than argued about.
| Approach | What it does | The catch | Score |
|---|---|---|---|
| Conversational interview | A guided text conversation that collects what a form cannot ask | Survey evidence said only about half of consumers would take a virtual assistant over a person | 82.5 |
| Bank data | Cash flow pulled from banking APIs to build a fuller picture | Backward looking, and every competitor was building it | 78.8 |
| On chain attestations | Verified claims about a borrower without exposing the underlying data | Answers get coarser. Over eighteen rather than a date of birth | 78.5 |
| Document OCR | Machine reading of whatever the borrower can photograph | Thousands of form types to learn, and no differentiation at the end of it | 77.5 |
| Virtual reality broker | A broker in VR asking about the borrower's future | Hardware costs, no in-house experience, richer competitors | 71.3 |
The winner became a short PRD. Three features. A transition page so a decline turns into a next step instead of a dead end. A guided conversation that gathers what a drop down cannot. A forward looking estimate that scores the answers. The MVP was deliberately a Wizard of Oz, a human behind the conversation, because the hypothesis worth testing was whether borrowers would answer at all, not whether the model could be built. It shipped with an explicit not doing list, acceptance criteria per feature, and a two week measurement plan that moved false rejections from one in four to one in eight.
The matrix ranked the conversational interview first. The evidence in my own deck argued against it, and I picked it anyway because it fit what the company was good at. A scoring matrix records a judgment. It does not make one.
Doubling throughput by writing better tickets
Engineering output per developer, measured in story points, doubled over my time there. No new hires, no new tooling. The change was that stories arrived with the ambiguity already removed, including the acceptance criteria and the parts we had decided not to build.
This is the least glamorous claim on the page and the one I would defend hardest. Most of what looks like slow engineering is a specification that forces an engineer to guess, then rework the guess. Writing the story properly is a product job, and it is cheaper than any other way of buying velocity.
The other side of the balance sheet
Originating quickly only matters if there is capital waiting on the other end. I built an investment product backed by mortgage assets, using on chain attestations so an investor could verify what the asset was and what lien sat behind it without the borrower's personal information riding along.
The privacy property is the interesting part and it is the same trade the qualification work ran into. An attestation returns a verified claim rather than a document. It answers the question the counterparty actually has, and refuses the rest of the file.
What I do not know
I left in March 2020, right before rates fell and refinance volume across the industry went vertical. Thirty-six hours at a startup's volume is a different claim from thirty-six hours at several times the applications, and I was not there to learn which one we had built.
The other thing I would change is the measure. Speed was the number the company repeated, and it was the right thing to chase, but elapsed time is a lagging indicator of something duller and more useful. How many times a human has to touch a file. If I ran this again that number would be on the wall next to the clock.