Stelios Stavroulakis: Learning What Not to Do, and Building What Comes Next

In Quilter’s Humans in the Loop series, certain conversations feel like they trace a straight line through time. Not because the path was easy, but because it was intentional from the very beginning. Stelios Stavroulakis is one of those rare people whose present work looks uncannily like what he set out to do decades ago. He did not stumble into engineering, nor drift toward artificial intelligence through trend or convenience. He pursued it deliberately, rigorously, and with a holistic view of how systems behave over time.

Stelios holds a PhD in Computer Science from the University of California, Irvine, where he focused on reinforcement learning and multi-agent collaboration. Before that, he studied electrical engineering and computer science at the National Technical University of Athens, the top engineering school in Greece. At Quilter, he brings those two disciplines together to tackle one of the hardest problems in modern hardware design: how to explore enormous combinatorial spaces while respecting the unforgiving constraints of physical systems.

Credentials alone do not explain why Stelios fits Quilter so precisely. What matters more is how he thinks, how he learns, and how he frames complexity as something to be explored rather than avoided.

‍

Origins: Engineering as a Foregone Conclusion

For Stelios, engineering was a natural progression shaped by his curiosity and surrounding environment. “I always knew I was going to be an engineer,” he says. “I have videos of me when I was six years old speaking to the camera and saying, ‘I want to become a scientist when I grow up.’” That early certainty manifested as a quiet confidence that allowed him to go deep rather than wide, to commit fully without wondering what else he might have been.

As a young child, his primary language was spatial, making geometry an early, tactile experience. He remembers carrying a small right-angle ruler with him, drafting intersecting lines just for the pleasure of it. "I loved the underlying rules of space, and I could see them everywhere", he recalls. "Even in art class, I was fascinated by vanishing points, using converging lines to translate a 3D world onto a 2D canvas."

At the same time, growing up with a father who taught control theory as a university professor and older brothers pursuing electrical engineering, conversations about systems and feedback were the standard background noise at home. Long before he knew the formal math of system stability, he was already absorbing the intuitive rules of balance, trade-offs, and logic through natural osmosis.

That pure joy of discovery never left him. For Stelios, engineering is not just a professional role he inhabits, it is simply how he engages with the world.

‍

Journeys in Engineering: From Signals to Learning Systems

At the National Technical University of Athens, Stelios pursued a five-year diploma equivalent to a master’s degree, with exposure to analog and digital electronics. He built PCBs the old-fashioned way, using photolithography processes that required patience and precision. That hands-on grounding in physical artifacts mattered. It taught him that design is never abstract for long. Eventually, electrons have to move.

At the same time, a lifelong fascination with music pulled him toward signal processing. But standard analytical tools had their limits for complex problems like polyphonic music transcription. A summer workshop at Stanford on music information retrieval introduced him to machine learning, and something clicked: signals were no longer just to be analyzed. Given enough examples, a system could also learn to recognize patterns on its own.

But learning from existing data was only half the puzzle. Stelios wanted to build systems that didn't just observe the world, but generated their own experience by acting within it. His diploma thesis applied reinforcement learning to solving the Rubik's Cube, an early indication of his attraction to sequential decision-making problems where brute force fails and strategy emerges over time.

That thread carried him to UC Irvine, where he immersed himself in reinforcement learning research and completed five research internships during his PhD. Each explored a different vertical, expanding the areas in which reinforcement learning could be applied.

At the Pacific Northwest National Laboratory, he worked on reinforcement learning for industrial power control. At Zebra Technologies - a company he describes as “the biggest company you've never heard of because they are the creators of the barcode that everyone uses but nobody knows about” - he encountered the complexity of multi-agent reinforcement learning. Warehouse scheduling exposed him to non-stationary environments, where "all your convergence guarantees go out the window," and success depends on coordinating many rational agents with limited local objectives.

Later, at Qualcomm, Stelios worked on large language models, applying reinforcement learning to issues like temperature control to reduce hallucinations. He was also selected from over 400 applicants for the Reinforcement Learning Open Source Festival, collaborating with Microsoft researchers on optimizing context injection for language models. Across these experiences, a common theme persisted: learning systems only work when you respect the structure of the environment they live in.

Why Quilter: Constraints as Opportunity

After years of academic research and industry experimentation, Quilter felt inevitable. “It just felt like a match from heaven,” Stelios says, pointing to the rare convergence of his electrical engineering background and his reinforcement learning expertise.

At Quilter, he is energized by the core challenge: solving a profoundly difficult combinatorial problem. Placement and routing are one expression of this, but for him, the problem is universal. “These combinatorial problems, they’re all really difficult. There is no clear solution. And the way you approach and model your problem is mostly what is going to define your success.”

This perspective is rooted in the reality of the domain. “There is no optimal board,” Stelios explains. “Translating schematic intent into physical hardware looks more like a negotiation of trade-offs.” The real question is how to navigate among many viable solutions and decide which trade-offs matter.

This philosophy extends to how learning systems improve. Reflecting on Quilter’s work, he notes, “When you want to learn something, you have to learn what to not do as well as what to do.” Constraints are not obstacles. They are the boundaries, as well as opportunities, that make superhuman performance feasible.

Looking ahead, Stelios is excited by the prospect of unifying placement and routing as “two players in the same game,” bringing them closer together rather than treating them as separate stages. It is an ambition rooted not in spectacle, but in coherence: fewer disjoint rules, more integrated reasoning.

Culture and Collaboration: Learning Together

Despite his deep technical focus, Stelios did not expect to join a startup. What stood out to him most about Quilter was the culture. “I have been blown away with the amount of care,” he says, emphasizing how trust and attentiveness shape daily work. With a small team, collaboration becomes personal. Ideas circulate freely, and everyone is heard. “I learn so much from this exchange of ideas that we have on a regular basis with the team.”

That openness mirrors his broader worldview. He delights in spotting hidden parallels between technical systems and everyday life, noticing, for instance, how a computer reclaiming unused memory resembles a legal system clearing outdated regulations, or how a parking lot filling up with awkward gaps mirrors the way a hard drive fragments over time. "Something extremely satisfying lights up in my brain when I map two things that seem disconnected," he says. It is the same impulse that drives innovation in complex systems: pattern recognition across boundaries.

Beyond the Workbench: Structure, Chaos, and Play

Outside of Quilter, Stelios's interests echo his professional philosophy. He plays footvolley, a Brazilian fusion of beach soccer and volleyball that demands creativity within strict rules. He loves salsa dancing, especially improvisation. "If it were to be ultimately chaotic, it wouldn't be fun," he observes. "If it would be ultimately structured, would it be fun?" The joy lives in the balance. "Whether you are looking at the layout of a city, a piece of music, or even hardware design," he reflects, "the beauty always lies in the middle. It is all about finding out how much freedom you actually have within your constraints."

Food and coffee connect him back to Greece. His ideal meal is simple and personal: Greek souvlaki, paired with a Greek iced coffee. The ritual matters as much as the taste. It is comfort, culture, and memory in one.

A Line to Remember

To distill his thinking, Stelios quotes Harry Klopf: "Other learning paradigms are about minimization; reinforcement learning is about maximization." This is more than just flipping a sign. Minimization simply conforms to the known, while maximization actively explores the unknown, driven by a curiosity to discover what else is out there. This mindset of continuous exploration is the engine of true novelty, and it perfectly captures the approach at Quilter.

‍