00:00I'm Allison LaForgia with HousingWire, and this is Housing Stack On Demand, powered by HousingWire and the BasisPoint.
00:06We're bringing technology showcased at the Gathering by HousingWire directly to you,
00:10giving lenders the insights they need to make smarter, more strategic tech stack decisions with the latest innovations.
00:16This demo features Inferd.
00:18Inferd is a Silicon Valley-based AI company specializing in intelligent document processing and contextual document understanding.
00:26They help mortgage enterprises automate high-variability document workflows at scale.
00:33Inferd's proprietary AI models are built to handle complex, semi-structured, and unstructured documents beyond traditional OCR capabilities.
00:41Walk you through an end-to-end loan journey using Inferd's Mortgage Check AI and Ally.
00:47Let me start with this loan package.
00:49This is what a typical loan package looks like.
00:52A mix of pay stubs, applications, loan application, bank statements, sometimes clean, often messy, all bundled together.
01:02And this is where every loan typically begins, borrower documents coming in, in all shapes and forms.
01:10Here is a user console of our Mortgage Check AI solution.
01:14What I'll show you is how we can take this loan package and turn it into structured, validated, and decision
01:22-ready data.
01:24So let's start by uploading a loan package.
01:27Typically in a real setup, these documents come through either a POS, LOS, or APIs, but the flow remains the
01:37same.
01:37We upload the file, and that's where the processing begins.
01:42The first step is the system automatically identifies different document types that you see on the left here.
01:50It splits that combined file, fixes orientation and scan quality, if applicable.
01:57So there's no manual prep, no sorting that your users have to do.
02:01The system takes care of all of that automatically.
02:04This step is very critical because every downstream system depends on clean data.
02:10So under classification review, everything comes in one big messy file.
02:16The platform separates it and classifies it into individual documents automatically.
02:23So if you see here, there's page number three to seven was classified as a bank statement.
02:32If it was incorrect, the users have an option to change it to the correct document type.
02:38The users also have an option to move pages around.
02:42So if this page number seven was part of a previous group, they can simply drag and drop these pages
02:49and correct the classification review of our platform.
02:53Every user interaction that you do is considered for retraining purposes.
02:59And every action that you do comes with a confidence score.
03:04Anything below the confidence score can be reviewed by the end users.
03:09But most of it flows straight through.
03:13So this essentially becomes more of an exception flow rather than a manual step.
03:21So after classification review, the system automatically extracts all the borrowers from the loan package.
03:32Once classified, it maps the names across documents and groups, everything correctly.
03:39Jane Doe was identified as the only borrower in this loan.
03:43So essentially, you don't have to manually figure out which document belongs where.
03:49The system automatically takes care of the association.
03:54So the system here is saying Jane Doe provided a credit report, pay stub, a verification of employment, and then
04:03W-2.
04:05After this step, the system moves on to data extraction.
04:10So once we classify and map the borrowers, we move into extraction.
04:17Here, the system captures all the key data points from pay stubs, bank statements, and other supporting documents as well.
04:27So let's click on one of the pay stubs on the left.
04:31The system will show you a preview of the pay stub, that single page that was uploaded.
04:37You can click on view on CUI button to review the data extraction.
04:44So if you click on the borrower name, our system is automatically going to show you where the borrower name
04:51was extracted from.
04:52It puts a green bounding box and a leading line indicator to make it easier for the user to navigate
04:59to the exact data point.
05:01And this level of detail is very important because this downstream system and downstream logic depends on it.
05:09So similarly, we extract data from all possible document types in the loan package.
05:17Like there are different bank statements.
05:21You can go inside each of these bank statements and can review all the transaction level information as well.
05:28So at the end of data extraction, what you're getting is structured data across the entire loan package, not just
05:38individual document types.
05:42After, during this process, the system also identifies if there were any documents that were missing.
05:52So it flags missing documents, inconsistent document uploads, or any data discrepancies like mismatched incomes or missing or conflicting values.
06:07All of this is surfaced in one place, instead of teams manually comparing documents.
06:14This is where a lot of time savings comes in.
06:20After we have reviewed the discrepancies and the missing documents, the users have an option to review how the final
06:31application and the set of documents were organized into a clean stacking order.
06:37So documents are grouped into their right sections.
06:41So if you see here, all the EARLAs and loan application related documents are under the application folder.
06:48All your bank statements, letter of employment, any earnest money receipts or purchase agreements, they are classified under the assets
06:59category.
06:59Same goes for credit, same goes for credit, employment income.
07:04The good part about this stacking order is this can be pushed directly into the LOS or it can be
07:12exported as a bookmark PDF ready for lenders to do their audits.
07:17The bookmark PDF will clearly show you all the correct grouping based on the stacking order that was defined by
07:27the users.
07:29This stacking order is configurable depending on different types of lender requirements, document requirements, and how you want the data
07:38to flow to the LOS system.
07:43Now let's go back to the user console.
07:46So at this point, what we have done is we have taken this messy loan package and turned it into
07:53structured, validated, and organized data.
07:57But the real challenge starts here, applying guidelines and making decisions.
08:02That's where our agentic solution, Ally, comes in.
08:08So let me click on the QC assistant on the right hand side of this loan.
08:16Ally is our agentic layer.
08:18Here it takes the structured data from Mortgage Check AI and applies all the mortgage underwriting lender specific logic automatically.
08:29So let's start with income.
08:31So this salaried agentic AI was automatically enabled.
08:38So let's click on the review button.
08:43Here the system will show you all the loan metadata, like the loan number, amount, the type of the loan
08:49at the very beginning.
08:53Ally validates any required documents, applies guideline rules.
08:58It can apply both Fannie Mae and Freddie Mac guidelines automatically.
09:09And it also has the capability to execute custom rules by uploading those rules in an Excel sheet.
09:19At the end of this process, this salary agent basically flags all the discrepancies that it found
09:28So the first one here is, is regular monthly earnings consistent year over year.
09:35So it automatically flags all of these issues wherever it's needed and shows everything that's already verified.
09:43If I scroll down, it gives you a snapshot into the current monthly income details of the borrower,
09:51starting with the pay period, start date, the pay period, end date, and what's their monthly regular income.
09:58More importantly, it captures the YTD monthly earnings.
10:02It also comes with a chatbot where you can ask some meaningful questions as it pertains to the context of
10:09this loan.
10:11So what Ally is doing is, it's preparing the file for underwriting decisions in the LOS.
10:18It doesn't make the final decision, it just enables users with all the required data points,
10:23so they can take the correct loan decision as it applies to the context of this loan.
10:31What's powerful here is transparency.
10:34Every number and decision is fully available and visible for the end user to take the right decision.
10:42Now let's go back to the QC review dashboard.
10:48Let's look at assets.
10:51So I'm going to click on the asset review review button.
10:56Now in case of assets, our agentic solution goes through all the bank statements, money, earnest receipts,
11:06and credit card statements, and it applies reserve requirements.
11:11Just to make sure, like are there any large deposits,
11:14are there any large deposits properly sourced and documented?
11:19It flags any large or unusual transactions.
11:23So if I click on suspicious transactions, it only shows me transactions which were above
11:30five thousand or about ten thousand dollars.
11:32So you can configure these thresholds depending on what you consider as a suspicious transaction.
11:40So instead of going line by line through each of the transaction and bank statements,
11:45the analysis and the heavy look is already done for you.
11:49All the users need to do is review all of the data points and take it forward.
11:57Let's go back to the user console.
11:59So the entire flow essentially works together.
12:03Mortgage Check AI.
12:05Mortgage Check AI structures the data.
12:07Ally applies the logic on top of it.
12:11And all of this can be integrated with LOS, POS, APIs, and aligns with systems like DU and LP.
12:19Now imagine walking to work.
12:21The documents are already sorted.
12:23The income is already calculated.
12:25The issues are already flagged.
12:27The checklist is complete.
12:29All that's left is judgment.
12:32That's the shift that we are trying to enable with Mortgage Check AI and Ally.
Comments