Life in adtech: A day in the life of StackAdapt’s Data Science team

Right now, somewhere in our stack, a model is deciding what ad impression to bid on, and it has milliseconds to do it. 

• It doesn’t get a logged-in user to identify. 
• It doesn’t get an immediate ground-truth label telling it whether its last bid was the right call. 
• It doesn’t get a held-out test set with the answers waiting. 

It gets a partial, noisy picture assembled from signals across the open internet, and it has to commit before the auction closes. This is the environment our researchers and engineers work in every day.

Meet StackAdapt’s Data Science team

Over the past 12 years, our Data Science team has grown to roughly 62 people, split between machine learning engineers and applied machine learning scientists. 

We asked three of them—Richard, Kuan, and Zeynep, who between them have spent almost two decades building at StackAdapt—to tell us what it’s really like on the inside: the problems that keep them up at night, the wins that make all the hard work worth it, and the culture that turns a genuinely difficult technical space into a place where people stay and grow for years.

What a real day in the life of the StackAdapt Data Science team looks like

No two days look quite the same, but here’s how Zeynep, a technical manager, describes hers:

“I start most mornings by checking our deployed models and ongoing experiments. Part of my job is owning the portfolio of systems the team has shipped, which means continuously balancing new bets against maintaining the ones already running. After that, the day is a mix of people and technical problems—unblocking a cross-team dependency, working through where someone wants to grow in a one-on-one, or digging into a design doc or a set of experiment results.”

That one description contains the two forces every data scientist at StackAdapt is constantly weighing: the pull to build something new, and the responsibility to keep what’s already shipped healthy. 

She also tries to keep an eye on where research and the industry are heading, so she can help prioritize the highest-impact problems rather than the loudest ones.

Contrary to popular belief, most of the job isn’t training models 

Richard breaks down where the hours actually go into three buckets:

1. Defining the problem: Pinning down what the problem actually is and checking whether a solution already exists before building a new one.
2. Preparing the data: Acquiring it, understanding it, and cleaning it.
3. Putting the model into production: Validating the output, building a reliable production pipeline, and monitoring performance as the underlying data drifts.

“The modeling is almost the easy part once all of that is in place,” Richard says. 

If you’re coming from a role or program where 90% of your time went into tuning architectures, this is the biggest mental shift to prepare for.

You’re never working alone

Work here follows a pipeline, and Zeynep describes it cleanly:

Engineering (data & infra)  →  Data Science (modeling)  →  Product (client needs)

• On one end, engineering teams own the data pipelines and infrastructure—good ML depends entirely on high-quality, consistent data and reliable, efficient computing resources to experiment on. 
• On the other end, product teams help translate client needs into problems that can actually be modeled. 
• In the middle is the rest of the data science team—reviewing approaches, debating tradeoffs, and sharing ideas. 

Kuan’s favorite part of the job is exactly this: the deep-dive discussions with colleagues to map out new, cross-functional project ideas.

The clearest example of this pipeline in action is Richard’s lookalikes project, which identifies users who resemble the customers who’ve already converted for a given advertiser. 

It isn’t powered by one hand-tuned model—it runs on an automated training pipeline that regularly retrains a model per advertiser. 

The genuinely hard research problem isn’t squeezing more accuracy out of any single model; it’s building a pipeline robust enough to consistently produce high-quality models across a huge range of advertisers spanning different industries and different goals, all running simultaneously. 

That’s systems thinking layered on top of modeling—the kind of problem that rewards people who like building things that scale, not just things that work once.

The problems that actually keep us interested

StackAdapt is an AI advertising and orchestration platform that spans both programmatic and owned channels. On the programmatic side, StackAdapt helps advertisers bid on and buy ad space across the open internet, rather than inside a walled garden, which shapes many of the interesting technical problems the team works on.

Working with sparse, noisy signals. As Richard explains it:

“We don’t have a deterministic, logged-in identity for targeting, and it’s harder to observe what our ads actually caused. So we have to extract signals from systematically incomplete observations. It’s a great space to work in: few established solutions exist, and none is provably optimal, which means there’s real room to push the frontier.”

Figuring out what “good” even means. Zeynep points to evaluation as the problem that’s deceptively harder than it looks:

“From the outside, ‘does the model work’ sounds like a single number. In practice, the hardest part of most of our projects is assessing whether our models are actually better. Offline metrics don’t always move the online ones, and a model can look great on average but fail in ways you care most about. Some of the most important work is figuring out the right thing to measure before we optimize anything.”

Rigor and business impact aren’t in tension, they’re sequenced. Zeynep’s framing: “we use rigor to work out what is actually true, and business impact to decide what is worth building.”

The failure mode she actively watches for is a team spending weeks polishing an elegant model that never ends up adding value. Rigor for truth, impact for prioritization—a useful distinction whether you’re a new grad or ten years in.

How research actually makes it into production

A lot of teams talk about “staying current with research.” Here, it’s a practiced habit, not a slide in the onboarding deck.

Zeynep recently helped lead an overhaul of the team’s performance prediction models, moving toward neural network architectures pulled directly from recommendation system literature. Her rule for borrowing outside ideas: “the key is to understand why an idea works and adapt it to our data and constraints,” not copy-paste an architecture because it worked on someone else’s benchmark.

Richard sees this as existential for a lean team:

“AI is accelerating the pace of research, [so] the useful lifespan of any given method is short. A lean data science team can’t rebuild everything from first principles; tracking external work—papers, conferences, open source—is what lets us pull in new methods early and turn them into capabilities for our customers. It’s a big part of how StackAdapt keeps its edge.”

This is also why the team shows up in person at industry conferences. 

Zeynep is especially drawn to workshops that bring together the types of people wrestling with the exact problems the StackAdapt Data Science team obsesses over day to day, and she’s curious to see how the field is revisiting older problems, like user representation, now that AI has changed the toolkit available to solve them.

She also serves as a peer reviewer for WWW, ACL, and EMNLP, which keeps her close to what’s coming next before it ever shows up on a roadmap.

On where the frontier is heading, Kuan and Richard are watching the same shift through different lenses. 

• Kuan is focused on agents moving from “generating content” to actually reasoning, planning, and executing complex workflows—the exact space his team is building in. 
• Richard is watching architectural attempts to break two current limits of LLMs: that they can only “think” by emitting tokens, and that they can’t fold what they learn back into their own weights. 

He points to early work like Sakana’s Continuous Thought Machine as a first gesture at internal deliberation, with continual learning—systems that genuinely internalize experience instead of just carrying more context around—as the bigger, still largely unsolved problem.

The culture, in their own words

Ask Kuan to describe the team, and he’ll give you five things without hesitating:

1. Continuous learning: People are constantly learning from each other and exploring new ideas and technologies.
2. Willingness to take on challenges: People step outside their comfort zone on purpose.
3. Winning together: Collaboration isn’t a value on a wall; people genuinely support each other’s success.
4. Strong ownership: Individuals are trusted to take responsibility and drive projects forward.
5. Comfortable making mistakes: Experimentation is encouraged, and mistakes are treated as learning, not failure.

Zeynep names the same underlying trait when asked what predicts success on the team: “strong bias for ownership, intellectual curiosity, and humility to be wrong in an experiment and move on quickly.”

Richard describes how those values show up in practice: “It’s a low-ego, high-ownership team. When something breaks, the instinct is to jump on the problem together rather than figure out whose fault it was. And when something needs doing, no one has to assign it—someone just picks it up and owns it.”

What it takes to get in

We asked what separates a good candidate from an exceptional one, and the answers were consistent enough to be worth taking seriously if you’re prepping to interview here:

• Interrogate the question, don’t just answer it. Exceptional candidates notice the assumptions baked into a problem, ask what data they would actually have, and reason about tradeoffs rather than reaching for the fanciest method.
• Understand before you solve. One of the most common mistakes is jumping to a solution before fully understanding the problem, often paired with under-preparing on fundamentals.
• Handle disagreement well. Beyond technical correctness, strong candidates stay curious under disagreement and let the data drive the discussion when possible, because much of the work involves navigating ambiguous problems with no single right answer.
• Build the dumbest model first. A strong first instinct is not to reach for the best model right away, but to start simple, understand where it works and where it doesn’t, and improve from there.

Why people choose to stay

We asked why talented data scientists build long careers here, and the answers came back to the same themes: meaningful ownership, challenging problems, and the people around them.

You get real range. Kuan highlights the ability to move across projects and domains without switching teams. That variety broadens your technical judgment and keeps the work fresh.

You own meaningful work early. When asked how much ownership data scientists get, Richard’s answer was simple: “As much as they are capable of.” New hires take on real projects and autonomy from the start—not months of shadow work.

Your work directly shapes the product. The team owns the models, optimizations, and AI systems that form the foundation of StackAdapt’s competitive advantage. You’re not building around the product; your work is part of it.

You solve hard problems with thoughtful collaborators. Richard points to challenging, unsolved problems and a supportive team as major reasons people stay: smart colleagues, room to grow, and the opportunity to take on increasing responsibility as the company scales.

These aren’t perks on a benefits page—they describe what the day-to-day experience actually feels like.

Ready to build together?

Our Data Science team has grown to over 60 people—machine learning engineers and applied machine learning scientists working side by side on problems that don’t have textbook answers yet. 

If sparse data, messy evaluation questions, and owning something from framing to production sound energizing rather than daunting, we’d love to talk.

Check out our open Data Science roles on the StackAdapt careers page and come build with us.