From Machine Learning to AI Agents: Stefano Rosa on Building the Future of Healthcare Automation

I like to remember that what really kick-started everything was reading Ray Kurzweil's How to Create a Mind. The book came out in 2012, but I picked it up in 2017. It planted this idea in my head that intelligence itself could be engineered, and that we were already on that path.

And people at DeepMind were already showing the first signs of this idea. There was DQN learning to play Atari games straight from raw pixels, and then, right around when I read the book, AlphaGo beat Go, a game everyone thought was decades away from falling to a machine.

So I started looking for where I could actually study it properly. Back then, the UK stood out; it had a handful of universities that offered AI as its own undergraduate degree, which was genuinely rare at the time. That was enough to make me pack up and move.

Research is about invention, finding the gaps in what we understand; entrepreneurship is about innovation, finding the gaps between people and the solutions they need

The experiences have taught me how to solve problems where the answer isn’t obvious, and the data is never perfect.

That’s been very useful as a founder. In healthcare, real-world workflows are messy. There are edge cases, broken processes, missing information, and systems that don’t always work as intended. I’ve learned not to design for the ideal case, but for what actually happens in practice.

At Serene, that means we spend a lot of time understanding where things can fail and building around those realities from the start. My background has made me comfortable moving between technical details and the bigger problem we’re trying to solve, and making decisions even when we don’t have perfect information.

The most valuable lesson was learning to explain any idea in three sentences or less

I’d say the biggest factors were perseverance, having a strong team, and being very focused on sales.

We also had a product that solved a real problem. We listened closely to customers, kept improving it, and made sure the product got better over time rather than standing still.

Customer loyalty was also a major driver of growth. Many customers stayed with us and kept buying, giving us a strong base to build on.

And honestly, we were just very persistent. With a team of eight, everyone had to contribute, move fast, and keep pushing even when things were difficult.

I love taking an idea in my head and turning it into something people can actually use, see, or experience. Especially with technology, there’s something exciting about creating a tool or product that works well, looks good, and makes someone say, “wow.”

Building also means caring about the craft. Learning the tools, practicing for years, and paying attention to small details that most people may never notice but that make the final result feel right.

Pre-PMF, you're in discovery mode. The goal is to generate noise. You do things even the hard way specifically because each one generates data you can use to triangulate toward what actually works. You need to generate noise to find patterns and figure out which direction to actually go.

Growth is the opposite mode entirely. Once you have that signal, the job shifts to systems — thinking in terms of variables, processes, and feedback loops you can actually tune. It becomes about relentless optimization instead of relentless experimentation.

I used to underweight sales, assuming a great product would just sell itself. But that is more the exception than the norm. Getting that wrong taught me that sales and product are equally important and go hand in hand. Product is about making the “impossible” possible, and sales is about informing what 'possible' should actually look like

People when asked about AI, think of ChatGPT. But there is a new type of AI coming. A type that can actually use existing software and complete work the way a person does.

This is especially important in healthcare, where systems are notoriously fragmented. So much work still happens across legacy systems, insurance portals, EHRs, spreadsheets, and other tools that don’t work well together. Today, people spend a huge amount of time manually moving information between these systems.

I believe we are just at the beginning of a new type of product, and at Serene, we’re building intelligent agents that can take on complex administrative workflows and actually complete the work. Not just drafting a letter of medical necessity or suggesting how to respond to a denied claim, but handling that last mile between systems: logging in, moving information, submitting the work, and following it through.

I hope the technologies and companies I build have a real, lasting impact. In 10 years, I want to look back and know that what my team and I built helped millions of people live better lives, either directly or by making other important technologies possible.