A watt should be read, not guessed

Preditoring started from a boring observation: most datacenter energy dashboards do not measure energy. They multiply CPU utilisation by a nameplate figure and present the product as a fact. The number moves when the workload moves, which makes it look alive, and it is wrong by 20–40% in either direction.

That matters because you cannot optimise against a number you cannot trust. If the baseline is a guess, the saving is a guess, and the finance team is right to ignore it. So the first thing we built was not a dashboard — it was the measurement chain: IPMI and Intel RAPL per host, smart PDU outlet watts overriding them wherever a real instrument exists, and an explicit label on anything that fell back to an estimate.

The second thing was the write path. Reading is safe and therefore easy to sell; writing is where the savings are and where the risk lives. Rather than ask for trust, we built the policy engine first: rate limits, single-active-per-target, time-of-day windows, setpoint guardrails, a state snapshot before every change and a seven-day rollback after it. The product will refuse its own suggestions when a guardrail says no.

We are a small engineering team working with operators in Tunisia and France. If you write to us, an engineer answers.

Measure, do not model

Where an instrument exists we read it, and where one does not we say so. Every figure in the product is labelled measured or estimated, and the estimated ones are excluded from savings totals.

Act, do not alert

An alert transfers work to a human who is already busy. An action that is gated, logged and reversible does the work and leaves an audit trail behind it.

Reversible by default

Every write snapshots the prior state first. If we cannot describe how to undo something, we do not ship the ability to do it.

Come and check the claim

Point an agent at one rack for 90 days. If the measured saving is not there, you will have found that out for free.