MetisTwin

A living model of your operation.

Your assets, their limits and physics, and the world around them. Complete enough to test decisions before they touch the real thing.

F
system.environmentreplayable

world.weather

Weather field

world.market

ERCOT RTM

site.meter

POI meter

twin.graph

TwinGraph

site.bess

BESS unit

fmi: lfp cycle model

agent

Agent

physics: fmi + modelicasemantics: agent-readabletraces: recordedreplay: 90d+

TwinGraph

Not a 3D picture of your plant. A model software can work with.

Most twins are pictures. TwinGraph is a typed, versioned description that forecasting models, physics simulations, and AI agents all read from. It is an open format that outlives any vendor, including us.

Parisi-Labs/twingraph

Typed, executable, versioned decision-twin IR. Open source.

ercot-bess-01.twinv3
# twin: ercot-bess-01 · marcy 345 kV hub
node site.bess.unit_04: Battery {
chemistry = "LFP"
power_max = 50 MW
energy_max = 200 MWh
round_trip_eff = 0.937
physics = fmi("lfp_cycle.fmu")
}
edge site.bess grid.poi_meter : PowerFlow
edge market.ercot_rtm twin : PriceSignal
policy evening_dispatch {
horizon = 24h
objective = maximize(margin)
respect = [soc_limits, warranty, interconnect]
}

Physical modeling

Every node can carry the physics that make the decision real.

A battery has a chemistry, a degradation curve, and limits unlike the battery at the next site. The twin knows, so a cycle is valued the way that unit wears.

Battery unit

cycle degradation, round-trip efficiency, SOC limits

Solar array

irradiance response, inverter clipping, availability

Grid node

constraints, congestion, curtailment, market settlement

Flexible load

duty cycle, interruption cost, thermal state

Already maintain FMI or Modelica models? They attach directly, and decades of engineering work joins the twin.

Reward environment

Where decisions prove themselves, and where the next models come from.

Decisions prove themselves here, and the record of every test trains the next models.

Test before control

Candidate calls replay against your history. Only the ones that win move toward running for real.

Every test becomes training data

Every forecast, test, and outcome is recorded, and those records train the models underneath the agent.

Built for machine learning on graphs

The twin is one connected graph, so models can learn how a change in one place ripples through the rest.

Readable by AI agents

Every asset carries a plain-language description, so AI agents can walk the twin and assemble what a decision needs.

Connected twins

A site twin can become a fleet twin, then a market-aware operating model.

A battery joins a portfolio, a portfolio joins a market. String enough together and you approach a working model of the whole organization.

joinsjoinstwin.siteOne assetbattery, plant, or sitetwin.portfolioA portfoliofleets, sites, business unitstwin.marketThe market around itgrid, prices, weather, the worldworld data

Model the system before you automate the decision.

The twin comes first. Every call proves itself there before it runs anywhere real.