Inside Net AI: A Conversation with Orlando E. Martínez-Durive on AI, Mobile Networks, and Critical Infrastructure
Net AI is proud to have a dedicated team of diverse professionals working from around the world. To start the year off, we thought we’d share a conversation between Mariam Kalandia, our marketing assistant and Orlando E. Martínez-Durive, a Machine Learning Researcher at Net AI and a post-doctoral researcher at IMDEA Networks Institute.
In this interview, Mariam speaks with Orlando about his work at the intersection of academic research and real-world deployment. Based in Madrid, Orlando shares his perspective on AI for critical infrastructure, the challenges of deploying intelligent systems in mobile networks, and how initiatives like Open RAN, digital twins, and Safeguarded AI, the project Net AI is undertaking backed by the UK Advanced Research and Invention Agency (ARIA), can help bridge the gap between cutting-edge research and reliable, large-scale industry adoption.
From Academia to Net AI
Mariam: Thanks for joining me today, Orlando. Before we dive in, can you tell me where you are based currently, and also, how’s the weather?
Orlando: So I’m based in Madrid, Spain. The weather is a little bit cold right now, which I like. Most of the time, it’s very nice. It’s chill, but not too cold.
Something that I really like a lot about the weather in Spain is that even if sometimes it’s hot, it’s still very dry. That’s something I enjoy compared to places where it’s hot and the humidity is also very high.
Mariam: Marco (CTO and co-founder of Net AI) is also based in Madrid. Do you see him often?
Orlando: Yeah, I see him very often because I’m also working with him as a postdoc in his group. So we have this other space now where I’m more focused on academic research with him.
We collaborate a lot and have a lot of work that we want to do together. Worst case, I see him twice per week. Other times, I see him three or four times per week.
Mariam: And is that how you first came across Net AI?
Orlando: Kind of. I don’t remember exactly when, but at some point, the name popped up somewhere. Maybe it was while talking to Marco or through his LinkedIn posts. And because it was quite a catchy name, I wanted to know more about it.
So I didn’t ask him directly. I believe I went to social media first and then the website to get a sense of what it was. I liked it, and we’ll probably talk about this later, but it’s this idea of taking something theoretical, like an academic paper, and trying to build something based on that.
Then I found out that Paul (Net AI’s CEO) was also a PhD student at IMDEA, actually the first one. That gave me a sense of connection. He came from the same place where I was doing my PhD, so it developed a sense of familiarity with someone I didn’t actually know yet.
I also checked out a few of his papers. Then there was an event where people from IMDEA Networks, including alumni, came together and gave talks about what they’re doing now, their experiences, and how they see the future. When he was giving his talk, it aligned a lot with what I want to accomplish and how I look at the future.
So that was my first real touch with Net AI. After that, I remember thinking, “Hey, I should keep an eye on these guys. This seems interesting.”
Life at Net AI: A Startup Mindset
Mariam: What is your current role within Net AI?
Orlando: I’m a Machine Learning Researcher. But because we’re a small company with a startup spirit, we need to wear different hats at the same time. I really like what I do as a machine learning researcher, but I also like being involved in data analysis, processing large datasets, collecting data from open sources, and mixing them. I enjoy taking care of different things. At the same time, thinking about how to present our product to customers in the best way possible. That means developing nice interfaces and making sure the demo clearly represents the potential of our technology.
These tasks sometimes fall outside what a typical machine learning researcher’s role might be. But they’re very important, because you can have amazing technology with great capabilities, but if you don’t convey those capabilities in a simple way to a customer, you’re not going to make it.
Mariam: For a non-technical audience, which is myself, to be fair, could you explain the difference between an ML Engineer and an ML Researcher? I interviewed Sam (ML Engineer at Net AI) a couple of months ago, and he gave his perspective from an engineer’s point of view.
Orlando: In an ideal scenario, machine learning researchers are people who try to push the boundaries of knowledge. They come up with new ideas on how models and training can be optimized, and how to build smarter and more powerful models.
At the same time, they assess the complexities behind the models, identify bias, identify issues, and make sure the model is doing what it’s supposed to do and is aligned with the task. That requires a lot of theory, and people with academic backgrounds may relate to it more, though it’s not exclusive to them.
On the other hand, machine learning engineers focus on how to deploy these models into solutions. They make sure latency is as low as possible, the model is in the right format, and they handle the technical issues that come up when deploying something in production.
That’s the theoretical distinction. In reality, the lines are blurry. Sam also takes care of tasks outside a typical ML engineer role — he builds models, analyzes research papers, and ensures they’re as advanced as possible. And I sometimes handle engineering tasks as well.
Mariam: Is there a particular project or research area you’re especially excited about right now?
Orlando: Yes, I have a couple, and they all fall under the big umbrella of AI for critical infrastructure.
There are many things within critical infrastructure, but I specifically care about networks. Most of the time, we see networks as a way to watch YouTube, scroll Instagram, or message family. But the same infrastructure is used to reach out to emergency services.
It’s extremely important that this infrastructure is reliable. On the one hand, the network needs to work perfectly. On the other hand, we have AI models, agentic AI approaches, that promise to change how we approach technology, but at the moment, they’re not always fully reliable. People don’t fully trust them in critical scenarios.
I like to think that we are working toward a future where we can fully deploy AI agents and trust them to be reliable at all times, especially during emergencies.
Why Moving Carefully Matters
Mariam: There’s definitely some reluctance toward adopting AI agents into networks, which is understandable since it’s still a new concept. In your opinion, is this reluctance justified, or should we move forward already?
Orlando: I think the reluctance is necessary, and you can look at it from different angles.
One angle is regulation. Network operators are subject to government regulations and standards, especially regarding emergency services. If a network goes down for a long period, people can’t reach emergency services, and operators can face huge fines. When they receive a license, they agree that their network will support emergency communications.
Another angle is the quality of experience. The mobile communications landscape is extremely competitive. While only a few major operators own infrastructure, many intermediary companies, like virtual operators, resell access at lower prices.
These intermediaries earn revenue by acting as a middle layer, meaning major operators earn less per customer while covering the infrastructure costs. So operators need to ensure their service quality is as high as possible to avoid losing customers.
It’s also expensive to train models. If you want the best ones, you often need partnerships, which take time to form. Doing it alone requires massive investment in data centres, training, and data collection. Operators need to be confident they’ll see returns on that investment.
Mariam: So basically, they’re being asked to commit before knowing what the results will look like.
Orlando: Yes, exactly. You need to commit a lot of resources before having a clear idea of what you’ll get in return.
Research, Energy, and Network Intelligence
Mariam: Tell me about your work at IMDEA. You’ve finished your PhD, right?
Orlando: Yes, I finished my PhD recently, which makes me very happy. I’m currently working as a postdoctoral researcher at IMDEA Networks. I think I’m in a very good position, combining my role as a machine learning researcher at Net AI with my postdoc.
As a postdoc, I have time to think about new ideas that leverage mobile networks as large-scale sensors and to improve their operation. As an ML Researcher, I develop and implement state-of-the-art solutions to improve networks while addressing the practical aspects that academic papers often overlook. This last point is something that I always try to keep in mind during my research. Is what I’m doing practical and can be deployed within the current infrastructure, or is it more theoretical and futuristic?
Mariam: Can you talk a bit about your research?
Orlando: My research focuses on three topics:
Mobile network metadata analysis. I look at questions like: what services do people use most on their phones? Does that relate to socio-economic variables like income, unemployment, and even political orientation? Is 5G being adopted across all segments of the population, and how fast? What is the impact of major events like protests on mobile networks?
Routing and network measurements. It surprises me how fragile and resilient the internet is at the same time. On the surface, everything seems to work smoothly, but under the hood, it is a bit chaotic. A single misconfiguration can impact a huge portion of the internet. We see this when cloud services go down, and half the internet is affected.
Sustainability and energy efficiency. This is something I’m very passionate about, given that RAN deployments are expected to be denser and the need for Edge AI is growing. I have my personal concerns about the future, energy use, and CO₂ emissions.
Mariam: How does this research relate to Net AI?
Orlando: There’s a lot of academic work showing that mobile networks consume a lot of energy. Operators need to keep antennas running all the time, which increases operational costs.
My research examines what is currently being done to save energy and what more could be done.
At Net AI, we believe that there’s a clear gap between what is being done and what’s possible. One specific project we’re working on, funded by ARIA, is building a digital twin to measure and compare energy-saving approaches that build on AI.
Operators usually test new solutions through small pilots, which are costly and slow. Our digital twin lowers that barrier, allowing AI engines to be tested safely in an environment that closely replicates real networks.
We’re putting a lot of effort into making the digital twin as realistic as possible. I’m very excited about this because it can be a game-changer. It allows AI models to learn through feedback loops and reinforcement learning, improving with each iteration.
In this case, we’re building the “game” for the AI to play.
Mariam: Yes, it’s part of ARIA’s larger programme called Safeguarded AI, which connects closely to what we discussed earlier. It provides a level of assurance that energy-saving solutions will work and remain safe in large-scale network deployments, lowering the risk for MNOs to adopt them.
Orlando: Exactly. Operators can configure the digital twin based on their own networks, embedding their real infrastructure parameters. That gives them confidence that what the AI does in the twin will match real-world deployment.
The Future of Intelligent Networks
Mariam: Besides the digital twin project, what else are you working on at Net AI?
Orlando: I’m also working on collaborations with major companies. These projects involve making our models compatible with Open RAN and wrapping our capabilities into rApps and xApps.
Mariam: From your perspective, where do you see the biggest opportunities for ML research in Open RAN?
Orlando: Open RAN de-aggregates network components to avoid vendor lock-in and increase competition. One key component is the Non-Real-Time RIC, where rApps can be deployed without very tight latency constraints.
This makes it much easier to deploy intelligence in networks. Previously, implementing ML required big changes in the network. Now, the architecture clearly defines where data comes from and how actions are applied.
The biggest opportunity is this clear framework, which reduces implementation complexity and makes operators more confident when deploying ML solutions.
Mariam: Is the network easier to manage under Open RAN?
Orlando: The short answer is yes and no. There are advantages, but also challenges. Using components from different vendors introduces security and integration concerns. One weak component can compromise the entire network.
So while there’s a lot of improvement, it needs to be handled very carefully.
Mariam: Some Open RAN vendors have scaled back recently. Does that affect how ML fits into the future of networks?
Orlando: I don’t think it changes the big picture. Once people see what’s possible with AI models, they’ll find a way to implement them — Open RAN or otherwise.
The question isn’t whether it will happen, but how. There’s no way back.
And Finally,
Mariam: Okay, my final question is: What’s the most recent book, movie, or piece of media that’s made an impression on you?
Orlando: I have two, actually.
One is a book I read years ago that I have been thinking about a lot recently: The Diamond Age by Neal Stephenson. It’s about a dystopian future where a young girl gains access to a personalized AI device that helps her learn, grow, and navigate life. I read it as a teenager, and it felt like science fiction. Now we’re getting very close to that reality, which still blows my mind.
The second is a YouTube video called “We’re Not Ready for Superintelligence”, based on the ‘AI 2027’ essay. It discusses the exponential growth of AI and how AI systems might start improving themselves. We’re already seeing early signs of this.
It raises concerns about safety, bias, and governance. When things move this fast, we need to be extra careful. That video really resonated with me, because things can change very quickly, and it’s important that we understand what we’re building and do it safely.
Through his work at Net AI and IMDEA, Orlando highlights the importance of building AI systems that are not only powerful but also reliable and ready to be deployed in the real world. His perspective reflects a broader shift in the industry toward responsible innovation, where advanced research, practical implementation, and critical infrastructure needs must align to create lasting impact.
