Diagnostics that never leave the plant
Industrial AI diagnostics built to run entirely on the customer's own premises. Plant telemetry is read, monitored and explained inside the building it came from, so an operator gets fault diagnosis without sending operational data to anyone, ourselves included.
- DomainWind, solar and industrial plant operations
- EngagementOwn research programme, measurement first
- StageActive programme under development, not a shipped product
Privacy decided the shape of the build
Most industrial AI asks a plant to send its telemetry somewhere else. That single request is what stalls the deal, because operational data describes how a plant runs, what it produces and where it is weak, and no operator wants that crossing the fence for a vendor's convenience. This programme starts from the opposite constraint: everything the diagnosis needs, the retrieval, the monitoring and the model that writes the explanation, sits on hardware the customer already owns and controls.
The work is a research programme. An edge co-processor programme, designed to sit beside a standard industrial PC, carries high-throughput vector retrieval and real-time drift and anomaly monitoring. A local inference stack runs open-weight models, quantized for commodity hardware and fine-tuned for industrial root-cause analysis. A data engine turns real fault-labelled wind and solar telemetry archives, together with physics-grounded synthetic plant simulations, into training and evaluation worlds where the ground truth is verified. A research paper on the work is in preparation.
The plant cannot send its data away
What stops industrial AI at the gate is rarely the quality of the model. It is where the data has to go, what the model was trained on, and whether any of the claims were ever measured.
Operational data cannot cross the fence
Plant telemetry describes output, downtime and weak points in detail. Sending it to a vendor's infrastructure is a disclosure decision, and for many operators the conversation ends before the diagnosis is ever discussed.
A general model has never seen a gearbox
A model trained on the open internet can describe a bearing fault in the abstract. It has not learned how a particular drivetrain actually degrades in real telemetry, so its root-cause reasoning reads well and diagnoses badly.
Claims that were never measured
Diagnostic accuracy in this field is usually asserted and rarely tested, on data the buyer cannot see, with the failures left out. A claim that was never allowed to fail is not evidence.
Everything the diagnosis needs, inside the building
The pieces of the programme, each built so the plant keeps its data and every claim stays testable.
An edge co-processor beside the plant computer
A co-processor programme designed to sit next to a standard industrial PC and carry the work a diagnosis needs continuously: high-throughput vector retrieval, and drift and anomaly monitoring in real time, on telemetry that stays on site.
- High-throughput vector retrieval on the plant floor
- Drift and anomaly monitoring in real time
- Designed to sit beside a standard industrial PC
A local model that explains the fault
An inference stack of open-weight models, quantized to run on commodity hardware and fine-tuned for industrial root-cause analysis, so the reasoning that turns a signal into an explanation happens on the customer's own machine.
- Open-weight models, with no external service in the path
- Quantized for commodity hardware
- Fine-tuned for industrial root-cause analysis
A data engine that builds the worlds
Real fault-labelled wind and solar telemetry archives are combined with physics-grounded synthetic plant simulations into training and evaluation worlds where the ground truth is verified.
- Real fault-labelled wind and solar telemetry archives
- Physics-grounded synthetic plant simulations
- Verified ground truth in every evaluation world
Frozen gates before anything ships
Every capability claim is registered as an evaluation before it is measured, with the thresholds that decide pass or fail frozen in advance. A capability that does not clear its gate does not become a sentence on this page.
- Evaluations registered before measurement begins
- Pass and fail thresholds frozen in advance
- Unmeasured capabilities stay unclaimed
What keeps the claims honest
The habits that decide what this programme is allowed to say about itself.
Measurement before announcement
Nothing is described as working until a pre-registered evaluation says so. The evaluation is written first, the gates are frozen first, and the result is whatever the gate returns.
Streaming mathematics validated on real measurement data
The streaming monitoring mathematics behind the drift and anomaly work was validated against real quantum-hardware measurement data. That is a statement about the data the mathematics was tested on, not a claim that anything here is quantum.
A result that kills an idea still counts
Findings that close a direction enter the record beside the ones that open it. The approaches still standing are worth something because the discarded ones were written down too.
Nothing leaves the building
Retrieval, monitoring and inference all run on the customer's premises. There is no path in this design where operational telemetry is uploaded so that a diagnosis can come back.
What the programme can honestly say today
No accuracy figure appears here. This is an active programme under measurement-first development, and a number that has not cleared a frozen gate does not get printed.
Nothing leaves
Operational data
A plant gets fault diagnosis and continuous monitoring without a byte of operational telemetry leaving the building it was measured in.
Pre-registered
Every capability claim
Each claim is written as an evaluation with frozen gates before it is measured, so what ships is what passed its gate.
Research grade
The underlying work
The monitoring mathematics was validated against real data from demanding instruments, and a research paper on the work is in preparation.
Have data that cannot leave your site?
If the reason your plant has no AI diagnostics is that nobody will let the telemetry out, that is exactly the constraint this programme was built around. Tell us what you are trying to diagnose.
Start a conversation