Smart Twin — a digital twin with an AI agent built in
Industrial plants model initiatives, costs and emissions to 2030 — and an AI agent explains, changes and compares simulations in plain language.
The problem
Moeve’s industrial clients had to decide how to decarbonise plants that run on boilers, turbines and cogeneration — choices worth millions, made across spreadsheets that plant managers, energy engineers and finance directors each read differently.
Smart Twin had to put the plant, the initiatives and their economics in one model. And because that model would be dense by nature, it needed a second way in: one where people could simply ask.
Two ways into the same model
I led the design from the first definition to the version running in production — objectives, flows and prototypes validated with stakeholders, developers, accessibility, QA, the data analyst and the PM. The product works on two levels that share one source of truth:
- The twin — the plant drawn as a diagram of equipment and energy flows, with every KPI opening into baseline-versus-scenario detail.
- The agent — an AI assistant that lives inside the simulation, reads the same data, and acts on it when the user agrees.
The AI agent
The hardest part of Smart Twin was not building a simulation — it was reading one. Instead of adding more charts, I designed an agent that sits next to the model and works on four jobs, defined with the team from real questions users asked.
01 · Explain the output
A simulation can change a dozen assets over five years. The agent turns it into a story: what runs when, what gets replaced, and why the numbers move.
02 · “What if” — change the model in words
Users change assumptions in plain language — set this month’s heat demand, lower the investment limit. The agent restates the changes, asks for confirmation, runs the simulation and links to the results. Nothing is changed silently.
03 · Ask the data
Precise questions get precise answers, without opening tables: how much natural gas one asset consumed in a given month, which energy mix a year relies on.
04 · Compare with the master
Every new scenario is judged against the approved one. The agent lays out the trade-off — emissions, revenues, investment, demand coverage — in a single reading.
Designed for when it fails
An agent that acts on business data has to fail honestly. Errors are explicit, recoverable and never disguised as answers.
First to ship the system’s AI components
Moeve’s design system had defined conversational components on paper, but they had never been applied to a real product. Smart Twin was the first: our flows were where those components were tested, adjusted and validated in production.
More broadly, I built on the existing system instead of beside it — creating the new components the platform needed and running sessions with Moeve’s designers to avoid duplicates and agree a validation flow before anything entered the shared library. Accessibility was a release gate from the start, not a final check.
Outcome
