Technology
Three decisions shape every part of the platform. Everything below is a consequence of them.
Take in every signal at the resolution it was produced, from every altitude — fleet, site, inverter, string. Nothing gets summarised before it has been understood.
Learn the physics and the failure modes of the domain before modeling it. The specification of the asset is a better prior than a year of its operating history.
Act on the small number of findings that are worth acting on, with the evidence attached, ranked by what it costs to leave them alone.
The stack
Nothing exotic underneath. The advantage is in how the pieces are ordered and what each one is allowed to consume — not in the components themselves.
Before any of it works
Every manufacturer names measurements differently. Sungrow calls it MPPT 4 current, GoodWe calls it PV7 Current, Hopewind writes MPPT4 input current — all roughly meaning how much electricity is flowing in through one input.
That much is manageable. The real difficulty is that they also express the number differently, and the number itself gives you no clue which convention applies.
143 amps, 14.3 amps and 1.43 amps are all perfectly believable readings. A wrong number that looks obviously wrong is harmless — someone spots it. A wrong number that looks plausible flows quietly into performance reports, contract calculations and carbon claims for years.
Turn each manufacturer's private language into one shared vocabulary.
A canonical dictionary of 865 measurements across 62 asset classes. Vendor labels, register addresses, scaling conventions and unit quirks are resolved once, at onboarding, into one definition that every downstream consumer reads the same way.
Decide whether each reading can be trusted before anything consumes it.
Every reading is scored and labelled on arrival. Only readings marked ready are eligible to feed a model, a KPI or a report — and that is enforced in code rather than left to convention.
Give every measurement a permanent identity that survives the hardware.
An identity is built from the asset, the quantity and which one — never from the gateway, the connector or the vendor's label for it. Those change; the identity does not.
How we use AI
An AI helps match manufacturer labels to the shared dictionary during onboarding. By the time it is consulted, deterministic rules have already narrowed 865 possible meanings down to a shortlist — usually two, never more than ten. Its only job is to pick one and cite the evidence that convinced it.
Picks something that was not on the list
→ Discarded and logged as a fabrication
Cites evidence that does not exist in the document
→ Discarded
Sounds certain about a control setting that could switch equipment
→ Downgraded to needs-an-engineer
Top two candidates are too close to separate
→ Forced to don't know
Because the model can only ever choose from a list the system handed it, making something up is structurally impossible rather than merely unlikely. "Don't know" is a valid, expected answer — not a failure.
Nothing is ever mapped without an engineer reviewing the proposal, the evidence and the runner-up. And once approved, the decision is saved: at runtime the translation is a multiplication, with no model in the path.
AI cost scales with the number of distinct labels a manufacturer uses — never with the number of readings. That is the difference between a platform that is affordable at fleet scale and one that is not.
Deployment & integration
Read-only from the systems you already operate; findings pushed back into the tools your team already opens.
The default. Your telemetry, models, and alerts run in an isolated tenant on shared infrastructure.
Fit — Most operators. Fastest to deploy, lowest operational overhead on your side.
Dedicated infrastructure inside your cloud account or region, with the platform deployed into it.
Fit — Where data residency, procurement policy, or an internal security review requires it.
Local pre-processing at site where connectivity is intermittent or bandwidth is constrained, with modeling still centralised.
Fit — Remote assets and constrained links. Scoped per engagement rather than offered as a product.
Security & trust
Each operator's telemetry, models, and alerts run in an isolated tenant. Nothing is pooled and nothing crosses the boundary.
End to end, from ingestion through storage, using standard TLS and storage-level encryption.
Permissions are scoped by role and by site, so plant-level detail reaches the people accountable for it and no further.
Access and configuration changes are logged, so you can answer who saw what and when.
Multi-tenant cloud by default, with dedicated and in-VPC deployment where residency or policy requires it.
NDAs and standard enterprise data-handling terms are in place before any operational data is shared.
Mixed vendors, sparse instrumentation, an anomaly nobody can explain. The first call is with someone who can discuss the detail.