AI is advancing at an extraordinary pace but for many organisations, the challenge isn’t simply building models. It’s about turning breakthrough ideas into systems that work in the messy, complex reality of everyday business.
That’s what makes Dr. Petar Veličković’s story so powerful. As a Senior Staff Research Scientist at Google DeepMind and Affiliate Lecturer at the University of Cambridge, Petar’s work sits at the intersection of theory, experimentation, and impact. On a recent episode of Data & AI Mastery, hosted by Dr. Raoul-Gabriel Urma, Petar shared how foundational AI research becomes real-world technology, powering billions of experiences and even advancing scientific discovery itself.
Below are the key lessons for every leader seeking to bridge that same gap.
Innovation Starts with Curiosity — and Collaboration
Petar’s journey began not in a corporate lab, but in a high-school classroom in Belgrade. His early fascination with mathematics and computing grew into a lifelong passion for discovery. That curiosity, combined with a network of mentors and collaborators, ultimately brought him to Cambridge University, the Montreal Institute for Learning Algorithms, and later DeepMind.
What stands out is how Petar views his career as a “series of fortunate events” shaped by people; teachers, peers, and colleagues, who opened doors and challenged his thinking. It’s a reminder that innovation rarely happens in isolation. Whether in academia or industry, the most transformative breakthroughs emerge where technical excellence meets collaboration.
For organisations building AI capability, this underscores the value of culture: a shared curiosity and openness to learning that allows ideas to move from lab to product.
From Prototype to Production — The Hardest (and Most Rewarding) Step
One of Petar’s most celebrated projects is DeepMind’s graph machine-learning system for Google Maps, which helps predict travel times for billions of users worldwide. The model doesn’t just look at one route; it understands the topology of the entire road network, accounting for factors like congestion or road closures.
Yet as Petar notes, success didn’t come from a single paper or benchmark. It came from doing the hard work of turning research into a deployable system.
In research, a 2 percent improvement might be the headline result. In production, that 2 percent doesn’t matter if latency doubles or if the model can’t adapt to real-world change. DeepMind’s team found ways to make the system scalable, caching results, adjusting inference strategies, and balancing accuracy with responsiveness.
The result? A measurable reduction in “negative outcomes,” where estimated travel times were significantly off. Beyond individual users, this improvement rippled across industries relying on the Google Maps API, from logistics and rideshare to food delivery.
The lesson for data leaders is clear: real-world impact comes from engineering excellence, not just model performance. A successful AI project isn’t the one with the highest accuracy, it’s the one that works reliably for millions of people.
AI as a Partner in Discovery
Beyond applied projects, Petar’s research also explores how AI can accelerate scientific discovery. In collaboration with mathematicians from Oxford and Sydney, his team built an AI system that identified hidden structures within complex mathematical graphs, insights that led to new theoretical breakthroughs.
Here, the AI wasn’t replacing human creativity. It was acting as a thought partner, surfacing patterns that experts could interpret and build upon. This partnership between human intuition and machine pattern-recognition demonstrates AI’s potential to reshape research itself.
As Petar describes, “Mathematicians are no longer skeptical that AI can accelerate their work.” The goal is not automation for its own sake, but augmentation, using AI to expand the boundaries of what’s knowable.
Lessons for Every Data and AI Team
Across both stories, Google Maps and AI-assisted mathematics, Petar returns to a central theme: focus on the human outcome.
Too often, data teams stop at the point where the numbers look good. But as he puts it, “There is a gap between seeing one number being higher and actually building something people will want to use.”
Closing that gap means engaging with stakeholders early, understanding user needs, and thinking beyond academic metrics. It means designing systems that account for latency, scale, fairness, and sustainability, the real-world “non-functional” requirements that determine success.
For organisations adopting AI, this philosophy is transformative. Start with empathy for the end user. Build cross-functional teams that blend data expertise with operational insight. And never lose sight of the fact that AI succeeds when it makes life, or work, better for someone.
Looking Ahead — The Future of AI Scientists
Petar believes we’re moving toward a future where AI systems can autonomously explore, experiment, and even propose new research directions, what he calls “AI scientists.” These tools won’t replace human judgment, but they’ll accelerate it, helping researchers navigate vast possibilities faster and more creatively.
For business leaders, this points to an even greater imperative: to equip teams with the skills and mindset to collaborate with AI, not just use it.
Building Your Organisation’s AI Capability
At Cambridge Spark, we help organisations do exactly that, develop the technical and strategic fluency to harness AI responsibly and effectively. Whether through executive education, data-science training, or bespoke workforce development, we enable teams to move from exploration to impact.
As Petar’s journey shows, AI mastery isn’t about chasing numbers, it’s about curiosity, collaboration, and creating solutions that truly make a difference.
Learn how Cambridge Spark can help your organisation build the next generation of AI talent: CambridgeSpark.com
