Inside Net AI: A Conversation with Samuel Knight

23 Jul

In today’s fast-evolving telecom world, engineers like Samuel Knight are not just building smarter models, they’re shaping the infrastructure of tomorrow. In this conversation, I sat down with Sam, a Machine Learning Engineer at Net AI, to talk about what a typical day looks like in our small but ambitious team, the promise and friction of applying AI in telecom, and what keeps him motivated, both professionally and personally. Whether you’re in academia or in the telco industry, this conversation offers insights into where telecom is headed, how machine learning is evolving to meet its needs, and why culture still matters in tech.

Life as a remote ML Engineer

Mariam: Hello, Sam. Thank you for joining me today.

Sam: Thanks, Mariam. Happy to be here!

Mariam: Let’s start with the basics: where are you based currently, and what does a typical workday as an ML Engineer look like?

Sam: I’m based in Davis, California, a small university town near Sacramento, the state capital. It’s about two hours away from the Bay Area, which everyone knows about, but people don’t seem to know that much about Sacramento. It’s super nice, quiet.

Usually, my day starts at around six in the morning. I try to get some overlap with people finishing their workdays in the UK in case there are any blockers or things I need to address. I make coffee at seven normally, eat breakfast. Although timing my breakfast significantly depends on the imminent deadlines. So sometimes, I end up putting that off until I’m really starving, and then I have to rush out of my desk to scarf down some cereal.

My work as a machine learning engineer depends on the stage of the different projects we work on.

If we’re in more of an exploratory phase, then it’s usually reading papers and, working on some implementations of those papers, making sure that our models are up to date with the state of the art, or there are individual components from those papers that we can incorporate into our model without having to extract the whole thing.

It’s also doing code reviews. We have a pretty small team that works on the ML side, and we’re all pretty in tune with what the other is working on. It’s a high-contact environment, which is good.

If we’re in a more delivery phase, then most of my day can be taken up by looking at evaluation metrics from model training runs that we’ve done, making sure that we’re not seeing any big regressions in terms of performance.

And if we are seeing regressions, then trying to work out exactly what is going wrong during training. If it’s a data issue, for example, and we need to do some more feature selection or clean some features, etc.

Sometimes I work on more visualisation stuff, front-end components, and sometimes more back-end stuff as well.

Mariam: You said your work day depends on a project phase we’re in. Is there a phase you’d rather be working on, whether it’s exploratory or development?

Sam: I think the most exciting times are when we’re delivering something and we get to work as a whole team together collaboratively.

I think some of the research phases can be more siloed in the sense that you spend time researching, reading a specific paper, looking to implement it, and then going to review, whereas when we’re actually on the road to delivering a solution, it’s usually very collaborative, which is probably the most enjoyable time.

Inside Our Projects

Mariam: Are you working on multiple projects right now, or are you fully committed to one?

Sam: Right now, I’m focused on the ARIA project. I’ll leave the detailed explanation to Marco (Marco Fiore, our co-founder and CTO), but this project involves building generative models that simulate how network conditions change over time across an entire network.

Instead of predicting a single future value, like what throughput will be next week, our models aim to generate realistic sequences of data that closely resemble the actual patterns we see. In essence, we’re creating a simulator that can produce believable network data.

This is really valuable because accessing real network data is often difficult. It’s sensitive and expensive for companies to share or collect. So, if we can generate realistic data ourselves, we can use it to train and test models for other tasks like forecasting or anomaly detection. It helps us scale our work much more efficiently.

Mariam: So now you’re in the developmental phase with ARIA. Is it mostly collaborative or still solitary?

Sam: Yes, we’ve moved on a little bit from the more siloed, paper implementation phase, and now we’re all working on evaluating what parts of the model work, and what don’t. We’re looking at the evaluation metrics and trying to come up with a final stable model that we can then take on to the downstream tasks together. So that’s been really fun!

Mariam: Was there a type of project you were really passionate about in particular, not just right now, but in general during your time at Net AI?

Sam: Yes, I was really excited, and still am very excited about our EnergAIze solution and the forecasted models that underpin that.

We see this theme across the world and industries, of increasing need for energy efficiency, cutting costs, and running things smarter. Although I get the sense that the networking space is a little bit behind in that regard. And it makes sense — network operators are focused on delivering the best possible customer experience, and sometimes that feels at odds with energy savings.

Obviously, if you start tuning down power transformers and cell towers and then reducing the throughput they can accommodate, for example, then a customer might have their data needs not met, and that would be detrimental, and they might leave the network if they feel frustrated.

But the thing is, network usage patterns are super predictable. We see clear spikes and drops throughout the day and week. So there’s a real opportunity here. It feels like a missed one, honestly, if we keep running everything at full power just in case.

Mariam: You’re basically saying telcos shouldn’t be so worried about user experience, because there’s a way for them to both save energy and not sacrifice and maybe even improve the quality of experience?

Sam: The key thing is making sure that there’s enough capacity on the network for the users.

There are these clear periodic patterns where you get spikes in user demand for the network capacity and decreases in that demand, whether that’s day and night patterns or weekend versus weekdays.

But the “always on” approach of the past, or at least I think it will be of the past going into the next few years, optimises for making sure that those customers have the capacity needs met.

And industry has had a sort of a reactive approach in terms of meeting additional capacity of building all this extra infrastructure as demand increases. But, it’s still “always on” as the default.

Having forecasting models that are specifically optimised to stay above user demand while decreasing energy usage as much as possible, I think that’s the way of the future. I mean, not only do you save energy, but you cut costs, because you use less electricity. So, it makes sense from a financial perspective, as well.

And I think as these models get more and more sophisticated, it’s only going to make the case of achieving this balance better and better instead of putting all your eggs in one basket.

Why Telecom Is Slow to Trust AI and Why That’s Changing

Mariam: Is the fear of decreased quality of experience the only blocker for adopting a solution like EnergAIze, or are there other concerns telcos might have?

Sam: Good question, and just to preface this, as an engineer, I don’t get that much interaction with the actual customers. So everything that I’m saying is filtered down through Paul and Marco and things I hear.

I’d say one of the biggest barriers is simply that it’s a new kind of system. AI and ML aren’t new in general. Industries like banking have used them for years. Even in something as old-school as ATMs, there’s machine learning doing character recognition when you deposit a check. That’s been around for the last twenty years or so.

So, in some cases, it’s very embedded. In the networking space, I feel there’s been less adoption in general, less trust in these systems to deliver. It’s a huge industry and sort of slow-moving by nature, and there is a hesitancy to adopt these solutions from smaller companies like us, just from the trust factor and the scale.

I think it’s a matter of individual telcos being willing to trial some of these solutions on a small scale, maybe in a single city, or even a couple of antennas around the city, where you won’t have a huge customer impact as you’re tuning the system to the parameters that you want, and then rolling it out and building it out from there.

So yeah, trust is probably the biggest factor. But there’s still movement. We’re seeing some telcos start to explore this, even building their own internal pilots. I hope that momentum keeps growing.

Mariam: It does make sense. These are large organisations, and adopting something new often means stepping outside familiar territory.

Sam: Yes, and another aspect might be that we just recently entered this new era of 5G rollout, and already people are talking about 6G.

One of the newer capabilities with 5G is that you can start to manage individual cell towers via software. Tune them dynamically. That opens the door to applying ML at a much finer level. But that whole software-defined network management approach is still pretty new.

So in a way, it’s sort of like a double hurdle to overcome in legacy thinking. Not just applying ML models, but actually doing the custom management of the network, via software, in this fine-grained way. I think that’s another thing that the industry is still moving towards being comfortable doing.

Mariam: It’s kind of wild to think that 5G isn’t even fully rolled out, and yet 6G is already in development.

Sam: I think it’s one of these cases where these standards, for 5G, 6G… It’s this mashup of all these competing interests, all these massive mega networking corporations that are all trying to lay out how they want the standard to evolve based on the challenges they face and so on. It’s just a constantly changing landscape. So, it makes sense that as one standard crystallises, like 5G, and starts to be fully implemented, all these different stakeholders are thinking towards the future and how they want the standard to change and what sort of things they want out of it.

In one sense, it’s like a very slow-moving space with how big the corporations are. But in another, there’s constant activity behind the scenes.

Mariam: AI’s become such a buzzword lately, especially in telecom. It promises a lot and sometimes it delivers, but now we’re seeing governments stepping in with rules and ethical guidelines. Do you think that’s reasonable, or is it veering into moral panic?

Sam: I think there has been this huge intersection between the big advancements in generative AI and public concern. I get it. Trusting the information we see online is already tricky and when AI can generate fake photos, videos, or voices that feel real, that trust erodes even more.

So I understand the caution from that side.

But zooming in on what we do. Our models are non-LLM, non-text, very task-specific. They generate realistic network KPIs, or help forecast traffic. How much harm can someone do with that box versus a box that can make someone believe that you said something you didn’t?

So we’re not caught up in the panic. There’s not much moral ambiguity in simulating throughput spikes.

Where I do see LLMs showing up in telecom is in customer support tools or internal assistants to help manage network operations. And in that space, I think it’s actually empowering. It’s less about harm and more about productivity.

On Net AI’s team culture and working from abroad

Mariam: I want to talk a bit about our team and the way you interact with them while being so far away. How do you manage to stay connected?

Sam: I think it’s as simple as having regular meetings, touching base with both people in the ML team and the wider team in our regular weekly meetings.

I’ve also tried to push for more informal meetups, even if they aren’t work-focused. Since we’re remote-first unless you’re in Edinburgh, that human connection is something you have to consciously maintain. Talking regularly, even just casually, goes a long way.

Mariam: Since Edinburgh’s sort of our hub, do you feel your relationship with the team varies based on whether people are there or remote?

Sam: Not really. I remember when I was in Edinburgh and we had the old office, the work environment was a lot different. It just is when you’re in person working with other people.

There’s this very cool environment that you just sort of take for granted. But since I was pretty close with the Edinburgh people, when I moved away, I noticed that difference in the environment, but in terms of being close with the team, that hasn’t changed at all. So I’ve enjoyed that.

Mariam: Do you have any rituals or habits that help you focus while working from home?

Sam: I try to use timers — usually 30 to 45 minutes — to really lock in and work without distractions. Then I’ll take a short break: stretch, grab a coffee, go for a quick walk. I don’t do it as consistently as I’d like, but it helps.

And honestly, I start my day with a bit of mindless browsing, news, memes just letting my brain idle before switching into work mode. Everyone has their own system, but I think some kind of routine is key when you’re remote.

Mariam: I can’t believe I haven’t asked this yet, but how did you end up at Net AI?

Sam: I met Paul at university. I took his IoT course, and that was my fourth year. Then, in my fifth year, I demonstrated for that course. So I helped out with some of the tutorials. We stayed in touch, and after uni I spent a year in a full-stack software engineering role, working in advertising, not really my passion, but the team was great, so it was a great experience, but I was looking for a change.

Then Paul reached out about a company he was spinning out from his research, looking for engineers. I was immediately interested. I had a solid background in ML from university, so it wasn’t a complete leap, but it was still a big shift day to day. I interviewed and joined late 2021 or early 2022, I think? Feels like ages ago now.

Mariam: What’s it like having your former professor as your boss?

Sam: Surprisingly chill! He was already an authority figure, so the dynamic was familiar. Honestly, Paul’s easy to work with in both roles, and I think the whole team would say the same.

And the team that we’ve built here, I think, is of like-minded people who are super collaborative, friendly. I hope we can scale that up as we grow, but that’s one of the things I think we should hold on to. The type of people that we have here. It’s a good place to work.

The Personal Side

Mariam: What is something people would be surprised to learn about you?

Sam: Oh, that’s an interesting question. Something surprising to hear about me…

I don’t like saying it because I feel like it’s sort of cliché of anyone who’s done this will tell everyone they’ve done it, but I’ve been skydiving before, which was something I wasn’t comfortable doing. But I did it anyway. It’s totally not my thing. I’m not into adrenaline, but my girlfriend was super excited about it and convinced me. I was quietly enthusiastic too, but definitely terrified. One time was enough!

Mariam: That’s pretty cool! And it’s definitely on my list too. I was also surprised to find out you lived in Switzerland.

Sam: Oh, that is a good answer. I lived in Switzerland; I was born there. My mom’s from California, my dad’s from Scotland. They met while working in Switzerland, and my siblings and I all grew up there. I still go back often. My parents live there, so it holds a special place in my heart.

Media Recs & Nerd Outs

Mariam: Any books, podcasts, or shows you’ve been enjoying lately?

Sam: From the shows, even though I’m sort of late on the bandwagon, I started watching Succession, which is so much fun. It’s just very entertaining watching all these hyper wealthy people tear their hearts out. I think we just finished season one, so we’re making our way through slowly.

As for books, after watching Dune Part Two, I picked up the series again. I’d read the first one years ago, but now I’m working through Children of Dune. I’d forgotten how much I love sci-fi — it’s way easier for me to finish fiction than non-fiction. Research papers are another story… not quite as engaging!

I also discovered George Saunders recently. His short stories are incredible. I read three of his collections back-to-back and would recommend anything with his name on it.

Mariam: Those are some great recommendations. Thanks, Sam!

Sam: Thanks, Mariam. This was fun!

Connect with Us

Net AI is a network intelligence company building advanced machine learning solutions to help telecom operators run smarter, more efficient networks.

Samuel Knight is the Machine Learning Engineer at Net AI. He designs and trains models that are the cornerstone of our network optimisation solutions.

Mariam Kalandia is a Marketing Assistant at Net AI. She manages the social media page and led this interview.