Buying a home is not a quick checkout experience.
It is months of research. Spreadsheets. Trade-offs. Stress. Excitement. And usually the largest financial decision someone will ever make.
So what does it take to build technology in the middle of something that personal?
In our recent DevLab Q&A with Zillow, we spoke to Kelsey Juraschka, Principal Machine Learning Engineer, about what it really means to design AI systems people can trust when the stakes are high.
At Zillow, the focus is not on chasing trends. It is on building systems that can hold up over time in a space where accuracy, fairness, and clarity genuinely matter.
While many companies are still working out where AI fits, Zillow has been building with machine learning for nearly two decades.
As Kelsey puts it:
"Everyone knows about the Zestimate. It’s actually been around since 2006. So if the Zestimate is like Zillow’s child, it’s an adult. We’re not new to machine learning and AI here."
That longevity changes the starting point.
Teams already have experience running non-deterministic systems at scale. They have operated in a regulated industry for years. Evaluation, fairness, and long-term maintenance are not new conversations. They are part of the operating model.
When generative and agentic AI entered the picture, Zillow was not beginning from zero. It was extending a foundation that already existed.
Housing decisions carry weight. They are emotional, financial, and often tied to major life transitions.
As Kelsey explains:
"This is going to be the largest financial decision that anyone is going to make in their lives, right?"
That reality shapes how systems are designed.
It is not enough for a feature to function technically. The experience has to reduce uncertainty. It has to provide clarity. It has to support confidence.
Zillow’s investment in ethical AI, fair housing compliance, and robust evaluation reflects that responsibility. When you operate in a regulated market and serve millions of users, trust is built into the architecture from the start.
When over half of home buyers cry at some point in the journey, empathy becomes part of the product brief.
Rather than adding AI to isolated features, Zillow is working toward something broader.
"We’re trying to build one Zillow co-pilot that can really assist users all the way from dreaming all the way to keys in hand."
Home buying is not a single transaction. It stretches across search, affordability, mortgages, listings, and more. It can take months. Sometimes years.
Designing a system that spans all of that means coordinating expertise across multiple domains while still delivering a cohesive experience to the user.
The current focus on agentic AI is about connecting systems, teams, and workflows in a way that feels unified rather than fragmented.
Generative systems behave differently from traditional software.
Outputs can vary. Small changes can ripple. Testing requires more than running a single unit test and moving on.
Here’s how Kelsey explains the shift:
"Now we have to figure out how to test these non-deterministic systems. … We don’t know that it will work 10 out of 10 times. It might work nine out of 10 times. And we want to be able to quantify that within our evaluations."
In other words, one successful run is not enough. Teams need ways to measure consistency, understand failure rates, and build confidence through structured evaluation rather than assumption.
The real work happens below the surface, in infrastructure, patterns, and systems that can scale across an organisation without creating bottlenecks.
The tools themselves are evolving just as quickly as the products.
Zillow has experimented with AI-assisted development for years. Some tools were early and imperfect. Others are now embedded in daily workflows.
As Kelsey says:
"I think the thing with AI is you need to keep trying because something that doesn’t work today is going to work really well, probably in a month, three months, a year."
Prototypes can now be built faster. Ideas can be tested earlier. Feedback loops are shorter.
At the same time, the quality bar has not moved. Code still ships to millions of users. Review processes still matter. Collaboration across product, design, and engineering remains essential.
Speed has increased. Standards remain high.
The thread running through all of this is ownership.
Engineers are shaping architecture, defining evaluation patterns, collaborating closely with product and design, and contributing to systems that span the organisation.
Working in this environment means thinking beyond a single feature, getting comfortable with ambiguity, and caring deeply about how systems behave in the real world.
It also means building alongside people who take trust, fairness, and long-term impact seriously. If you want to explore how Zillow is approaching agentic AI, evaluation, and large-scale system design in more detail, watch the full conversation below.