Featuring:

• Nhan Van Nguyen, Head of AI
• Andreas Waal, Head of Engineering
• Erik Åsberg, Chief Technology Officer
• Maria Cervantes Keel, Senior Marketing Manager

The launch of AI Studio by eSmart Systems marks a meaningful shift in how we bring our technology to market. For over a decade, we’ve built computer vision AI for the power sector behind the scenes of Grid Vision, embedded in the workflows of more than 75 utilities worldwide. AI Studio is the moment we open that up, giving users access to the intelligence layer that powers our technology. I sat down with our Head of AI, Nhan Van Nguyen, and Head of Engineering, Andreas Waal, and our CTO Erik Åsberg, to talk about what we’ve built, why now, and what we’re already seeing from the people using it.

Maria: Nhan, let’s start with the basics. AI Studio – what is it, in plain terms?

Nhan Van Nguyen:
So excited that you asked. AI Studio is a full-stack computer vision AI platform that enables utilities and technology companies to create, explore and deploy production-ready models in minutes. It’s powered by our patent-pending Adaptive AI. In layman’s terms the platform that makes all of the AI capabilities that we’ve built at eSmart over the past decade accessible to customers as a service — through a web interface and APIs. There are three core components:

• Model Garden: explore the computer vision models we’ve already built, ready to use or customize for your specific asset types and inspection standards.
• Model Builder: create something net new from your own examples, without a labeling pipeline or training run. What used to require weeks of ML infrastructure, you can now do in minutes.
• Pipeline Builder: chain models together into end-to-end workflows – classify something, measure it, and route the results into your operational systems in sequence.

And all of it is API-first; natively ready for agentic workflows, so it integrates directly into whatever systems or AI agents you’re already running.

Maria: Andreas, from a product perspective: what problem were you solving?

Andreas Waal:
The honest version? We kept seeing the same friction. A utility would want to customize a model; say, adapting a defect classifier to match their internal standards rather than a generic definition, and the traditional process was just too heavy. You’d need labeled data, an ML pipeline, training time, and validation cycles.

Weeks, sometimes months, would pass before you’d see results. And if the definition changed, which it always does, you’d start the cycle over. That’s not how domain knowledge works in real operations. We wanted to build something that matches the pace of how people actually think and work..

Maria: What makes AI Studio genuinely different? There are plenty of computer vision tools out there.

Nhan:
At it’s core is what we call our patent-pending Adaptive AI technology. The key difference from standard fine-tuning or few-shot learning is that it’s not a one-off process. You provide a handful of example images, the model generates immediately. No training run, no annotation pipeline, and then it keeps improving as you give it feedback. When a definition changes, you update the examples and the model updates with it, in real time. There’s no retraining cycle to redo. That’s genuinely new.

The other thing that makes it different is what sits underneath it. We use foundation models as the feature extraction layer, the best available at any point in time. When a new, more capable foundation model comes out, our Adaptive AI inherits those gains automatically without any re-labeling or engineering effort. We benchmarked this directly: switching to a newer feature extractor model last year pushed our F1 accuracy on a crossarm classification task from 0.72 to 0.89 in under a minute. No additional data. No work. The model just got better because the world got better.

Erik Åsberg:
And for the user, that’s the part that matters. Traditional AI development asks you to commit a lot upfront: data collection, labeling, infrastructure, etc. before you know if it’s going to work. AI Studio inverts that. You can start with only a few images and a working model, then improve from there. It changes the risk calculus entirely.

Maria: So why open this up now to the public? eSmart Systems has had these capabilities internally for a while.

Erik:
Honestly, a few things came together. The underlying technology reached a point where we felt confident putting it in users’ hands directly; the Adaptive AI architecture is mature enough that it doesn’t need us as an intermediary. At the same time, we started seeing the AI landscape shift toward agentic workflows, where software isn’t built and deployed statically but assembled dynamically by AI agents. That world runs on APIs. And we realized that if we stayed inside a traditional SaaS model, we’d be limiting the future impact we could have as a product company. Opening our capabilities as modular, API-first services is how we stay relevant in the next era of enterprise software, not just the current one.

Andreas:
There’s also a practical pull from the market. The utilities and technology companies we talk to aren’t waiting for a perfect enterprise solution before they start experimenting. They want to try things, build something small, see if it works. The Free tier is a direct response to that. It lowers the floor enough that someone can go from zero to a working model in an afternoon. And they can confidently scale from there with AI Studio.

Maria: Who are you seeing use it now, and how?

Andreas:
The audience is broader than I originally expected. Utilities are the core, and the use cases there are what you’d predict; customizing inspection models to match their own asset definitions, building classifiers for defect conditions that are specific to their territory or equipment base. But we’re also seeing interest from technology companies that want to embed computer vision into their own products or platforms without building an ML team. And from inspection service providers who want to differentiate their offering. The API-first architecture makes AI Studio something you can build on top of, not just use.

Nhan:
What I find interesting is how quickly people get to the multi-model pipeline use case. Initially, you’d think someone picks one model and runs it. But the Pipeline Builder has resonated also. Classify something, measure it, chain the results into a workflow. That end-to-end composition is where the real value starts to show up operationally.

Maria: What’s surprised you? Things you didn’t predict when you were building it.

Andreas:
The speed at which people get creative with non-utility applications. We designed AI Studio with grid inspection as the primary context, but the architecture is domain-agnostic, and users figured that out fast. We’ve seen early interest from people applying it to use cases well outside of power; environmental monitoring, industrial inspection, things we’d never scoped, including even classifying flower species in images. That’s genuinely exciting. And it validates the approach: build something flexible at the architecture level and don’t try to constrain the use case too tightly.

Erik:
For me, the moment that stood out was watching someone build a working model in a product demo, from scratch, no prior examples, in about twelve minutes, and have it performing in a meaningful way by the end of the session. Opening it to the world, and seeing what users are creating in real time keeps me excited about what we’ve actually built.

Maria: Where are you excited to take it next?

Nhan:
The biggest thing I’m watching is how AI Studio fits into agentic workflows – where AI agents are calling our capabilities directly as tools, not as a platform a human logs into. We’ve designed the API layer with that in mind from the start. What I want to see is AI Studio becoming the computer vision reasoning layer inside broader, multi-agent systems. A planning agent that delegates visual inspection tasks to an eSmart Systems’ model and gets structured results back. That’s where this gets genuinely interesting at scale.

Erik:
From a product side, the near-term focus is on reducing every point of friction in the experience – from signup to first working model to first API call. The Pro tier we’re planning will open up the Model Garden to a broader audience at a lower commitment point than Enterprise, which I think unlocks a whole segment of users who are serious outside of big enterprise use cases. And then the longer arc is about making AI Studio the tool that anyone building visual AI in an industrial context defaults to, not just because of the technology, but because the experience and performance is genuinely better than the alternatives.

Maria: Last one – for someone reading this who’s been sitting on the sideline on AI, utility team, technology company, inspection professional, what would you tell them?

Andreas:
Try it. The Free tier is exactly what it says – no commitment, no sales process, no ML team required. Upload some images, build a model, see what it does. The best way to understand what AI Studio changes is to use it. You won’t regret trying it.

Nhan:
And don’t assume the constraint is data. The barrier has never really been data volume, it’s been the overhead of the process around the data. That’s what Adaptive AI removes. If you have a use case in mind and a handful of examples, that’s enough to start.

Want to build your first model?

Sign up for the AI Studio Free tier at ai.esmartsystems.com – no ML expertise, no labeling pipeline, no training cycles required. Start with a handful of images and have a working model in minutes.

Ready to go further? Reach out to our team about Enterprise access.