Real-World Impact Through AI/ML Transformation

Rocket Mortgage Canada (RMC) is a digital-first mortgage broker helping Canadians access competitive home financing options. Unlike its U.S. parent company, which is a lender, RMC operates as a brokerage—serving as a bridge between borrowers and lenders. RMC is responsible for gathering, reviewing, and preparing the required documentation for lender-side loan officers to process applications efficiently and in full compliance with Canadian financial regulations.

To deliver a faster and more streamlined borrower experience, RMC sought to automate its internal document workflows—particularly around the manual and time-consuming process of reviewing and assembling complex applicant financial records.

To meet regulatory requirements and assist loan officers, RMC must analyze and package loan applications with extreme precision. A large part of this process involves reviewing applicants' financial statements—checking, savings, RRSPs, investments, and credit cards—from a wide range of issuers including major banks, credit unions, and fintechs like Wealthsimple.

Applicants frequently submit inconsistent documentation, including screenshots from phones, mobile apps, and scanned printouts. This inconsistency—along with differing layouts, formats, and poor scan quality—makes it extremely difficult to extract and analyze financial data accurately.

RMC's team had to manually trace income sources, verify asset levels, flag suspicious transactions, and generate summary and index pages linking each transaction to the correct document and page. This work typically took more than a day per application and slowed down the loan approval process significantly.

Softmax Data partnered with RMC to build a custom end-to-end AI pipeline that automates financial document analysis, compliance checks, and summary generation. The system begins by ingesting PDFs and classifying them by type, issuer, and account.

We then built custom models to extract transaction tables, standardize all formats into a unified structure, and normalize columns across a wide range of document layouts. This ensured every entry was tagged with consistent fields: Date, Description, Deposit, Withdrawal, and Balance.

A second layer of the system aggregates all transactions into a single financial timeline, allowing the model to map fund transfers between accounts, detect income consistency, and calculate asset summaries as of any given date.

To enable intelligent inquiry, we built a Retrieval-Augmented Generation (RAG) agent capable of answering natural language questions like “What is the total asset as of March 31st, 2025?” or “Has income from employer X been consistent over the last 6 months?”

Our solution also auto-generates a complete PDF package, including an index page linking transactions to their source documents, a summary sheet highlighting flagged items, and a neatly stacked bundle for the loan officer to process quickly and with confidence.

Finally, the system includes an ML deployment pipeline that automatically retrains the model on newly annotated data, ensuring it continually adapts to new formats and improves accuracy over time.

The impact was dramatic. What once took more than a day of manual work per application now takes under two minutes, allowing the team at Rocket Mortgage Canada to scale operations without scaling headcount.

The RAG assistant empowered loan preparers and compliance staff to get precise answers instantly without combing through dozens of pages, while the document packaging pipeline delivered clean, lender-ready summaries and indexes with every file.

By deploying this end-to-end Document AI solution, RMC has increased throughput, reduced human error, improved compliance, and brought faster approvals to their borrowers—all powered by AI.

In 2019 Absolute Results (AR) helped car dealers sell over 93,000 cars. They did this by shifting dealerships from a reactive culture to a proactive “appointment culture.” Instead of salespeople waiting for customers to walk into the showroom, AR gives dealerships training and technology to actively schedule appointments with clients.

This growing category of sales is called Appointment Driven Private Sale Events—and AR is the global leader. It’s all driven by data. AR uses creative messaging and targeted data to sends millions of personalized event invitations to potential car buyers.

While AR was the leader in Appointment Driven Sales the landscape of their industry was rapidly changing. To maintain their position they had to keep improving the accuracy of their personalized messages.

They specifically wanted to improve which vehicles they would recommend as upgrade options. They needed data to help them predict which car John Smith would want when it was time to upgrade. That kind of accurate recommendation would drastically increase the conversion rate.

But how could they accurately predict that information for thousands of potential clients? People don’t buy new cars everyday. That limited how much AR knew about the buyers and their preferences. Even the number of transactions for most of the vehicle models was quite limited. Traditional methods rely on having huge amount of data for each car model and customer to make accurate predictions for the customer’s next purchase; this made it impossible to use traditional recommendation algorithms. To make things even more complex, any solution would have to accommodate a wide range of data qualities, siloed data sources, and CRM systems that weren’t used to talking to each other.

The CTO of AR, Josh Heppner heard about the powerful potential of Softmax Data’s machine learning and data science solutions. Could we teach computers to amass thousands of data points and interpret it so AR could sell more cars?

Softmax worked closely with the IT team at AR and executive team to design, develop, and implement a highly scalable custom-built recommendation engine. The engine leveraged both deep learning models and statistical models to predict individual consumer’s buying habits and the exact upgrade vehicles they would want.

It looked like magic but the powerfully accurate predictions were all based on our powerful data processing solutions. We looked at consumers’ past purchase history, demographics, and geographic information as well as external market trend data.

If John Smith had owned a Toyota Camry, then a Honda Accord, had a college degree, made over $75,000 a year and lived in Toronto, we could predict with scary accuracy the car he would want to upgrade to next (it was a Lexus, by the way).

The algorithms were designed to be flexible. They allowed for dealerships to input their first-hand relationship with their client base. This customization around their sales process and expertise made the algorithms even more accurate.

The impact was immediate. Softmax Data’s machine learning solution increased revenue, conversation rate, sales funnel velocity and sales team performance. Within one month of the start of the project, AR’s recommendation conversion jumped by an eye-popping 320%.

And it just kept getting better. As SoftMax’s machine learning solution received more and better data, the predictions kept improving.

People often say that “data is the new gold.” And it’s true. But for a business to benefit from the “gold” of data you have to know how to mine it and refine it. Softmax Data did just that for AR. We used 100% of AR customer data to make the predictions. The result was that each dealership sent messages to clients that were more personal, more accurate, and more effective. And this meant that more cars sold. AR’s revenue increased by an estimated 27% as a direct result of Softmax Data’s work.