What four years of building AI have taught us

Long before generative AI became part of everyday conversation, Myrspoven was using machine learning and physics-based models to control heating, cooling and ventilation in buildings. We spoke with Max from our AI team about how the technology has developed since 2022, what years of operation have taught us and why the algorithm is only one part of effective building AI.

Today, AI is often associated with chatbots and large language models. But the technology used to optimize a building solves a very different problem: how can a building anticipate what it will need, rather than simply react after conditions have changed?

This interview has been edited for length and clarity.

What were you trying to solve in 2022?

Max: Many buildings are heated and ventilated using static setpoints and schedules, or a simple relationship with outdoor temperature. These settings are often tuned once and rarely revisited.

But conditions change throughout the day. The sun may hit the façade in the afternoon, occupancy may be lower on a Friday, and weather and previous HVAC decisions all influence what happens next.

We wanted to move from reactive control to a system that could anticipate what the building would need next. To do that, it must understand how each building responds to changing conditions and adjust its decisions accordingly.

How did the technology develop?

Max: We began with established machine-learning methods, including different types of neural networks. They could fit the data, but they required a lot of it and were difficult to interpret. That lack of transparency becomes a challenge when a model's output is used to control heating and ventilation. You need to be able to check that a decision makes sense.

In spring 2022, we shifted towards learning a building's differential equations from its own operational data, using methods developed through open academic research. The basic idea is to combine machine learning with what we already know about physics.

Heat transfer follows well-understood physical principles. By giving the model relevant physical building blocks, it does not have to learn everything from scratch. Instead, it can determine how those building blocks fit together in each building.

Every building is different, but the underlying physics is not.

What could the system do by the end of 2022?

Max: By autumn 2022, we could learn a building's thermal dynamics from its historical data, predict indoor temperatures several hours ahead and optimize heating and ventilation setpoints against those predictions.

That first version established the foundation we needed. We had one model per building, a focused set of signals, manually tuned comfort constraints and protective code to support reliable control.

Much of the continued development focused on detecting frozen CO₂ sensors, handling missing signals and ensuring that an inaccurate prediction could not become an unsuitable setpoint. This showed us early on that the AI model is only one part of a dependable system.

What have several years of operation taught you?

Max: Three things stand out.

First, data quality is central. Sensors freeze, drift or get replaced. Meters are renamed and signals disappear. The system must recognize when data does not make sense and avoid acting on it. Detecting unusual signals also helps the people operating the building identify when something is not behaving as expected.

Second, buildings change. Maintenance is carried out, seasons turn and tenants move. A model that accurately represented a building at one point needs to keep adapting. This is why continuous retraining, stability monitoring and guardrails have become so important.

Third, trust is operational, not only statistical. The people responsible for a building need to understand what the system is doing. Interpretability, dashboards and visible reasoning therefore matter alongside accuracy. An AI system needs to earn enough trust to remain part of daily operations.

How has the system evolved since then?

Max: The philosophy remains the same: learn how the building behaves, then optimize decisions against that model. But almost every layer supporting it has been rebuilt.

We started with open-source research libraries and soon adapted the models to JAX, a numerical computing framework that allowed us to run calculations more efficiently. By 2024, we had developed our own modelling and training stack, with physics-informed feature libraries for different signal types. Temperature and CO₂ behave differently, for example, and the model structure reflects that.

The scope of control has also grown. The system now considers relationships between systems, components and signals across a building, while respecting comfort limits and operational boundaries. Models are retrained, data quality is monitored and new protective mechanisms are added as we learn from additional situations.

How is this different from generative AI?

Max: The current AI conversation is largely shaped by large language models. They are useful tools, including for working with code and testing ideas, but they are not the foundation of the AI that controls a building.

Our building AI models physical behaviour and uses optimization methods to calculate control decisions. It is solving a different kind of problem. AI is an umbrella term, and two products may use the same label while relying on very different technologies, inputs and decision-making processes.

What are you working towards now?

Max: Scale without dilution. The system is built to operate across portfolios, and we continuously improve the foundation to make each new connection more efficient, adapt faster to change and keep operation transparent as portfolios grow.

We are continuously strengthening the technical foundation and developing a new interface that will make it easier to monitor and manage AI performance across an entire portfolio. The focus is on making deployment and daily operation even more efficient as portfolios grow.

The differentiator is everything built around the algorithm: operational experience, data handling, guardrails and continuous adaptation.

My test for any building-AI claim is simple: how long has the system been operating in buildings, and what happens when something goes wrong?

That is where years of experience make the difference. The future of building AI will not be decided by who talks about AI the most, but by what the technology delivers inside buildings, decision after decision.