Building Real-Time ML Infrastructure: How Varant Zanoyan Turned Airbnb's Internal Tools into Zipline AI

The Problem Nobody Wants to Admit

Here’s something that doesn’t get talked about enough in the ML world: the models aren’t the hard part anymore. What’s actually killing your velocity is the data infrastructure underneath them.

Varant Zanoyan learned this the hard way while working at Airbnb, where he was embedded with machine learning teams tackling real problems at scale. The trust and safety team was dealing with fraud detection—and if you know anything about Airbnb, you know it’s essentially a payments company moving money across borders. That makes it a magnet for fraud attacks, and the pressure to catch bad actors in real-time was intense.

“Every reason for Airbnb in many ways is also a payments company,” Varant explains. “It moves money across bank accounts, across borders. And as a result, it’s the target of a lot of the same fraud attacks as traditional payment processors.”

The teams wanted to move faster. They wanted to use more real-time data to detect shifting user behavior and protect the platform better. But here’s where it got interesting: the bottleneck wasn’t the ML algorithms. It was the data pipelines.

Building the First Solution Inside Airbnb

When Varant and his team started digging into why these ML projects were moving so slowly, they identified something critical: the data infrastructure—the pipelines used for both training and serving models—was where the majority of complexity and friction lived.

So they built tools. And they were really good tools.

“We started building tools to see if we could take that iteration loop from months down to something better, and we eventually got it down from months to days,” Varant recalls. “That was a pretty big breakthrough.”

But here’s where it gets more interesting: this wasn’t just a fraud detection problem. The search and personalization team was facing similar challenges, just at a different scale. When someone searches on Airbnb, thousands of listings need to be ranked in real-time, personalized for that specific user. Every click triggers a new search, a new geobox, new listings to rank.

The infrastructure challenge was even more complex. But the core problem was the same: data pipelines were the bottleneck.

The Birth of Chronon (and Eventually Zipline AI)

What started as internal tooling at Airbnb eventually evolved into Chronon—an open-source data management engine that would power ML infrastructure. The adoption was organic and telling. Stripe, Netflix, OpenAI, and other major companies started using it. The open-source community embraced it because they were all facing the same fundamental problem.

That’s when Varant and his team made the leap: if this many companies need this, maybe there’s a standalone platform here.

Zipline AI was born from this realization.

What Makes Zipline Different: A Core Decision That Mattered

When building Zipline, Varant had to make a fundamental architectural choice that would define the entire product. And this is where most feature platforms get it wrong.

“I think the biggest one was right at the start,” Varant says when asked about core decisions. “And I think this is one thing that differentiates us from other pieces of infrastructure in this space, which is that we are going to tackle the computation problem.”

Here’s the fork in the road: other feature platforms use a “bring your own compute” model. Essentially, they say: “You handle computing your features. We’re a feature store plus some metadata. You figure out the rest.”

It’s tempting because it’s simpler to build. But it doesn’t actually solve the user’s core problem.

Zipline went the harder route. They decided to own the computation problem end-to-end. Why? Because Varant had spent years sitting next to the users building these systems. He understood their pain intimately.

“The downside is that it’s a much more difficult skill to tap,” he admits. “And saying, okay, this is the usability you guys handle your features, up to you, the rest of it—it’s tempting because it’s far simpler, but it really doesn’t solve the core of the user challenge.”

The real issue is that data infrastructure evolved in layers: Hadoop, then Spark, then Kafka, message buses, streaming systems. Users had to stitch all these tools together. There was no end-to-end platform.

Zipline decided to be that end-to-end platform.

From Incubation to Enterprise

One of the advantages Varant had was incubating Zipline inside Airbnb for five to six years before spinning it out as a standalone company. This meant he could make mistakes, completely rearchitect things, and run massive data migrations—all within a single company’s infrastructure.

“It was really nice to be able to add an area of being able to be an incubating ground for this thing to have real use cases at scale, and so to make mistakes,” he reflects. “That wouldn’t simply be there by the way. Every time we made a big change like that, we had to think about how to migrate these teams that were relying on us still. It’s far easier to do it there than if we were doing this whole thing as a startup.”

Now that Zipline is a standalone company, the roadmap has shifted. The core engine—open-sourced and battle-tested—is stable. The focus is on building the enterprise platform around it: integrations, access controls, multi-tenancy features, and all the things that make it work seamlessly across different organizations.

The Lesson for Technical Founders

There’s something valuable here for anyone building infrastructure: sometimes the right move is to build inside a larger organization first, validate the problem at scale, and then spin it out. It gives you the runway to make architectural decisions that would be nearly impossible to recover from as a young startup.

And sometimes, the harder technical decision—owning more of the problem rather than less—is actually the one that resonates with users.

Want to hear the full story of how Varant built Zipline AI from internal tooling to a platform solving real-time ML infrastructure challenges? Listen to the full episode where we dive deeper into the technical decisions, early customer conversations, and the journey from Airbnb to startup founder.

Check out Zipline AI and connect with Varant on LinkedIn.