Digital twin
A software component that mimics the behavior of any physical system or product.
Physical systems are closed, have hidden states and give partial, noisy and delayed information. The twin is open, transparent and estimates what you cannot measure.
By David Kadleček, PhD, Group CTO & co-founder · first published 2021, updated 2026Act before a problem appears.
The same method for a person, a production line, a building and a bank customer: collect what the system tells us, keep one live model of it, and see trouble coming while there is still time to change the outcome.
A digital twin of a patient
Every source of health data feeds one model of the person: the twin sees the trend long before any single test does, so care can step in while the risk is still only a risk.
An illustrative scenario
Catching it before it becomes a disease
- 1Signal
Over weeks the smartwatch shows a slowly rising resting heart rate and shorter sleep; in the chatbot the patient mentions tiredness; the last lab result is borderline.
- 2Prediction
None of it would raise an alarm alone. The twin combines it with the patient’s history and predicts a growing risk of a metabolic or cardiovascular problem.
- 3Intervention
The clinician gets an early warning with the reasons behind it. The app proposes a check-up and an adjusted diet and movement plan; the health coach follows up.
- 4Outcome
New measurements show whether the trend turns. The disease does not have to happen — and the twin keeps learning what works for this patient.
A digital twin of production
Machines, lines and quality checks become one live model. The twin knows the state of every machine, sees anomalies early and explains why they happen.
An illustrative scenario
Fixing the spindle before it ruins a batch
- 1Signal
Spindle vibration drifts up, still within tolerance. Quality inspection sees a few more borderline parts and the MES shows slightly longer cycles.
- 2Prediction
The twin reads the three signals together, matches them to past failures and flags an anomaly: bearing wear, with a likely failure within weeks.
- 3Intervention
Maintenance gets the task for the next planned stop, the planner moves urgent orders to another machine and quality tightens inspection meanwhile.
- 4Outcome
No unplanned downtime and no scrapped batch. The repair happens when it costs least.
A digital twin of a building
The 3D model, sensors and building systems are joined into one live building. The twin tests “what if” before it happens and runs the building for comfort, energy and maintenance.
An illustrative scenario
Getting through a heatwave for less
- 1Signal
The forecast shows a heatwave. Meeting rooms on the sunny side are fully booked and one chiller is drawing more power than it used to.
- 2Prediction
The twin simulates the next days on the BIM model: which zones will overheat, when the peak load comes and that the chiller is likely to fail under it.
- 3Intervention
The BMS pre-cools at night when energy is cheaper, bookings move to cooler zones and maintenance checks the chiller before the peak.
- 4Outcome
Comfort holds, the peak load drops and a failure is avoided. The same method scales from one building to a whole district.
A digital twin of a bank customer
Transactions, products and life events become one live picture of each customer. The twin sees a change in someone’s situation early and routes the right next step to the right person — or solves it automatically.
An illustrative scenario
Helping before a payment is missed
- 1Signal
A customer’s salary starts arriving irregularly, card spending shifts to basics and savings are being drawn down. Nothing is overdue yet.
- 2Prediction
The twin compares this with the history of all customers, explains the root cause and predicts a rising risk that loan payments will be missed in the coming months.
- 3Intervention
A task with the situation and the suggested solution lands in the private banker’s list; the app offers a payment holiday or a better-fitting credit product and budgeting tips.
- 4Outcome
The customer gets help when it still helps, the bank avoids a default — and a loyal client stays. The same twin spots life events for well-timed offers.
Watch the twin get sharper with every iteration.
Bearing wear cannot be measured directly. The twin estimates it from noisy, partial signals, and the more information it integrates, the narrower its uncertainty.
- Signals integrated
- –
- Observations used
- –
- Estimation error
- –
- Uncertainty now
- –
- Twin alert
- –
What the twin does that the machine can’t.
We put the physical system into its full context: its historical, present and future states, and all interacting systems. It is not realistic to have 100% information at the beginning, so we integrate it in iterations. The twin becomes more precise after each one.
Open and transparent
All actual and historical data through an API in a unified model.
Suppresses noise
Suppresses measurement noise and model errors.
Estimates hidden states
From historical states and actual observations of the physical system.
Makes predictions
Predicts the development of machine states, so you can act in advance.
Active, not passive
Generates alerts as a reaction to an important change, anomaly or state reached.
Partition tolerant
Provides states before they arrive, or while the machine is disconnected.
Mathematics first. Tools second.
Reduced-order modelling
A mathematical model much simpler than the real object that still keeps the behaviour that matters to you — far smaller, with a controlled loss of accuracy.
System identification
Find the model’s parameters from measured inputs and outputs — physics-based, data-driven or neural.
Data assimilation
Continuously update the model with measured data and any additional information — Kalman and particle filters.
Optimal control
A control loop that steers the real object while minimising your objective, e.g. control effort or large outliers.
Questions we hear first.
How is it different from a dashboard?
A dashboard shows what was measured. A twin also estimates what cannot be measured, suppresses noise, forecasts with uncertainty and keeps working when data stops arriving.
Do I need all my data from day one?
No. We integrate information in iterations, and the twin gets more precise after each one.
Is it a wrapper around ChatGPT?
No. The core is a mathematical model of your system trained on your data. Language models are one optional layer on top, for interpretation.
Digital Twin Readiness Check.
Start with one line or process and the data you already have. In one month you find out what its twin can see — before you invest in more.
You provide
- Access to up to 3 data sources (e.g. sensors, MES, quality records)
- Your current process or calculation
- Known pain points
You get
- Feasibility and accuracy read
- Prioritised use cases
- Proof-of-concept plan and business case
- Roadmap to scale to further lines and plants