Data
The world-state layer behind the runtime.
Metis connects to lakehouse-scale histories and builds a typed picture of real systems and the world around them: what was true, what changed, what actions were available, and what happened next.
Scale
Trillions of rows
Shape
State, action, outcome
Use
Train, inform, backtest
Data loop
Trillions of rows are only useful when they replay the decision.
Metis is not industry-specific at the data layer. It treats every operational system as a sequence of state, context, feasible actions, and outcomes. That shape lets us train models, inform the live runtime, and score counterfactuals against history.
Train the models
Large-scale histories teach the time-series models how real systems move, how regimes break, and which context matters before the next state changes.
Inform the runtime
Live telemetry, constraints, events, forecasts, and external state become the context Metis reads before it recommends an action.
Backtest the decision
Every recommendation can be replayed against historical baselines, actual operator behavior, and the outcomes that followed.
World state
The model needs more than the target series.
A forecast gets sharper when it can read the state surrounding the series. The data layer preserves that context, aligns it in time, and keeps it available to both training and live decisions.
System state
Telemetry, inventory, availability, load, location, SOC, alarms
World state
Weather, markets, outages, demand shocks, traffic, events, prices
Decision state
Constraints, contracts, candidate actions, risk limits, operator intent
Outcome state
Settlements, service levels, utilization, costs, failures, emissions
Lakehouse-native
We meet the data where it already lives.
Metis can run against existing lakehouses, warehouses, object stores, operational systems, and decision logs. The point is not to move everything into a new silo. The point is to make history usable for the models and the runtime.
Lakehouses
Metis reads
Delta, Iceberg, Parquet, warehouse tables, object storage
Metis returns
Feature sets, replay cohorts, training corpora
Operational systems
Metis reads
EMS, SCADA, ERP, TMS, MES, historians, custom APIs
Metis returns
Current state, limits, exceptions, action traces
External feeds
Metis reads
Weather, market data, public datasets, commercial feeds, event streams
Metis returns
Forecast context and regime labels
Decision logs
Metis reads
Bids, dispatches, schedules, overrides, tickets, playbooks
Metis returns
Counterfactual baselines and evaluation sets
Historical replay
Backtests become product evidence.
Before a runtime writes to an operating system, it should prove itself on the history the buyer already trusts. Metis packages historical state, candidate actions, and outcomes into replayable evaluation sets.
2019-07-18 14:00
high load
heat event
reserve
$41k avoided
2021-02-15 07:00
asset constrained
scarcity
hold
risk reduced
2023-09-04 18:00
fleet available
demand peak
dispatch
lift proven
2025-05-22 11:00
maintenance window
normal
defer
cost avoided
Bring the history your systems already trust.
Metis turns that history into training context, live runtime context, and the replay set that proves whether the decision would have been better.