The early office machinery at Valon included a piece of equipment rarely celebrated in fintech: a printer. Linda Du and her co-founders were trying to build modern software for mortgage servicing, a business where money moves, escrow balances change, rules multiply and a late document can become a serious problem. Yet before their system could automate the work, somebody had to do it. So they processed tickets. They mailed documents. They originated and serviced hundreds of loans by hand. In a trade addicted to the elegant demo, they began with paperwork.
It was an expensive form of education. Valon could not simply arrive with code and ask large, regulated institutions to trust it. The software needed to function inside the place where borrowers, investors and government agencies all meet. Du and her colleagues therefore built two businesses at once: a mortgage servicer and the operating system beneath it. One generated the problems. The other learned to solve them.
By August 2026, the apparent detour had reached its intended destination. Carrington Mortgage Services completed its acquisition of Valon Mortgage, an operation then servicing about 810,000 loans, and chose ValonOS as the core platform for a combined book expected to approach two million loans. Valon, the parent technology company, could now concentrate on software. The proof had been separated from the product, and the product survived.
The route into the machine
01 · A career of converging lines
The résumé makes sense in reverse
Du studied computer science and statistics at Harvard, then tested both halves of that education in unusually literal fashion. She interned as a software engineer at Google. She became an investment banking analyst at Goldman Sachs, a quantitative researcher at Two Sigma and an investor at Samlyn Capital. Code, capital, models, operations: the sequence looks eclectic until mortgage servicing enters the picture. Then every chapter finds a job.
Servicing begins after a loan closes. It includes collecting payments, managing escrow, communicating with investors, helping borrowers through trouble and obeying a library of rules whose footnotes have footnotes. The work is financial and computational, but also procedural and human. A system can be technically correct and still leave a homeowner staring at a bank account, wondering why the payment has not moved.
Du's public language tends toward systems. She talks about processes falling off a conveyor belt, controls embedded in workflows and a source of truth strong enough to support AI. But the desired outcome is simpler: fewer moments of needless uncertainty. She has compared the visibility of a mortgage payment with the tracking people already expect when ordering dinner or a car. If a delivery app can say where the noodles are, a mortgage portal ought to say whether the money has settled.
We always want our humans to be focused on the things that humans are really good at and for technology to take care of the rest.Linda Du, The Way Home podcast
02 · Operations as fieldwork
The edge cases were the curriculum
Valon's founding wager was that mortgage servicing was fundamentally a software problem. Its second wager was more interesting: nobody could solve that problem from the outside. The company became a servicer to expose ValonOS to actual payments, actual compliance clocks and the bizarre exceptions that arrive without an appointment. This was not customer discovery conducted over coffee. It was customer discovery with legal duties.
The arrangement also gave Du's team a blunt sales argument. They could show what had happened inside their own operation, where the platform had been tested under production pressure. A traditional vendor promises that the system will work. Valon had to live with the answer every morning.
The numbers show scale, but the useful detail is what happened between them. Valon's team learned how an agent can close a ticket while the underlying task remains undone. It learned how a borrower can disappear midway through a help process. It learned that a transfer between servicers creates a rush of questions precisely when the data is changing hands. Software had to notice the dropped thread, prompt the next action and preserve a record clear enough for another institution to inspect.
On a Freddie Mac podcast, the host noticed Du smile as soon as servicing transfers came up. It was the smile of someone hearing a familiar nuisance called by name. A transfer can leave a homeowner asking why a mortgage has been sold again, while the incoming servicer reconciles unfamiliar data and braces for more calls. Valon's rapid growth had made such transfers nearly continuous. Du's conclusion was not grand: communicate sooner. Give people information before anxiety turns into a phone queue. The observation is typical of her approach. Start with a sprawling institutional problem, then find the specific moment when a person becomes confused.
That emphasis helps explain Du's phrase “compliance by default.” In her telling, compliance should not depend entirely on a person remembering the correct move at the correct time. Where possible, the platform automates the step; elsewhere, it builds controls directly into the workflow. Regulation, in this model, is not a policy binder sitting beside the product. It is part of the product.
If you had unlimited people, what careful work would you ask them to do? Du uses that thought experiment to identify valuable automation: listen to every call, score every interaction and catch every stalled process.
03 · AI with a job description
Automation should make room for attention
Du is not shy about AI, but her examples are agreeably untheatrical. Valon has used language models to produce call summaries so an agent can listen instead of typing notes. It has explored reviewing every call rather than sampling a few. These are not parlor tricks. They are attempts to manufacture consistency in a business where missed context can become financial distress.
Her view also contains a warning. In a 2026 post, Du said Valon temporarily removed AI access for new non-engineering hires after noticing that the tool could let newcomers jump to plausible answers before they had built enough knowledge to judge them. Experienced employees accelerated; some new employees skipped the mental model. The measure was temporary and deliberately blunt. The goal, she wrote, was not less AI, but people with the judgment to use it well.
The episode reveals something about Du as an operator. She is willing to interrupt access to a fashionable tool when it interferes with apprenticeship, then ask the harder question of what a new employee must learn in an AI-rich workplace. Mortgage servicing rewards accumulated instinct: the ability to recognize which ordinary-looking detail will become tomorrow's exception. A model can return an answer instantly. A newcomer still needs enough experience to know whether that answer deserves a place in a regulated process. The glamorous technology, in other words, remains dependent on the unglamorous education.
That distinction is particularly sharp in servicing. A home is often a person's largest asset and, as Du has observed, their most emotional purchase. Efficiency matters, but so does the quality of the remaining human moment. When software takes the notes, the agent can listen. When a portal explains each step, the homeowner need not call simply to learn whether anything happened. Automation earns its keep by giving attention back.
Homes tend to be the most expensive purchase that someone makes and they also tend to be the most emotional purchase that someone makes.Linda Du, The Way Home podcast
04 · The company after the company
Selling the proof, keeping the thesis
Valon raised a $100 million Series C in October 2024, a handsome vote of confidence in a corner of finance better known for hold music than glamour. Yet the more revealing milestone came two years later. Carrington acquired Valon Mortgage and committed to ValonOS. The operating business that had served as the laboratory moved to a specialist with more than two decades in servicing. The lab equipment, in effect, became the commercial product.
The transaction did not remove Valon from mortgages. It clarified the role. Du described the plan in plain sequence: operate a servicer, prove the technology in production, then scale the technology beyond the original operation. Carrington's portfolio and Valon's existing loans were expected to bring the combined platform close to two million. The software would now encounter still more government servicing complexity, only as infrastructure rather than as the owner of the operating company.
There is a neat startup lesson available here, and it should be handled carefully. “Do the work yourself” is useful advice until it becomes an excuse to build three businesses. Valon's version worked because the operational burden answered a precise trust problem. Mortgage institutions did not merely need an attractive interface. They needed evidence that the system could endure the edge cases, reporting duties and consequences of the real job.
Du's achievement is therefore not that she made mortgage servicing look exciting. The more durable trick was refusing to require excitement. She and her colleagues chose a large, unfinished problem, entered through its least decorative door and stayed long enough to learn where the machinery catches. The printer was never the future. It was how they earned the right to build one.