AI for manufacturing that advises the floor and leaves the controls alone
Anomaly detection on machine data, a maintenance copilot that reads the manuals and raises the work order, visual inspection, supplier document extraction and production forecasting, for discrete and process manufacturers and the software vendors that serve them.
Unplanned downtime, scrap and paperwork are the three costs that AI reaches first. All of it works alongside your MES, CMMS and ERP, not instead of them.
Anomaly detection and predictive maintenance
Vibration, temperature, current and cycle data compared against each asset's learned normal; departures flagged with the signals that drove them. Touches historians, PLC data via the MES and the CMMS. Technicians confirm or dismiss, and the model learns. See anomaly detection.
Maintenance copilot
A technician asks about a fault code on a tablet and gets the manual page, the last three times it happened and a draft work order with parts, hands-free with voice where gloves are on. Touches the manuals library, the CMMS and the parts inventory.
Visual quality inspection
Camera-based detection of surface defects, missing components and assembly errors at line speed, with images and verdicts stored against the batch. Touches the line cameras and the quality module of the MES. See computer vision.
Supplier and specification document extraction
Certificates of conformity, material test reports, datasheets and supplier invoices read into the ERP and PLM with validation against the purchase order. Touches the ERP, PLM and the quality system. See document processing.
Demand and production forecasting
Order intake, seasonality and supplier lead times combined into forecasts the planner adjusts, with the assumptions visible. Touches the ERP and the planning tool. See predictive analytics.
Shop-floor knowledge assistant
SOPs, safety procedures, changeover instructions and shift notes searchable in plain language, in the operator's language, with the current revision only. Touches the document control system and the shift log. See knowledge management.
Compliance and data
Safety systems, OT boundaries and quality records
What we design for in manufacturing
Functional safety and machine control stay with certified, deterministic systems; AI advises and never actuates. The operational technology network is segregated, so AI services read through controlled gateways and nothing on the plant network depends on an internet call. Quality and traceability records must survive audits, so every AI verdict is stored with its inputs and version. Technical data may be subject to export control, which affects where models run. Worker data from wearables or cameras is personal data under GDPR and UK GDPR and needs consultation and minimisation. This is engineering guidance, not legal advice.
We run models on-premise or at the edge where needed, design the OT gateway with your controls engineers, and version every model, prompt and inspection threshold. The model supply chain security page covers how we vet the models themselves.
Reference architecture
Flagship workflow: anomaly alert to closed work order
Sensor and PLC data reach the AI services through a one-way gateway from the OT network. Anomaly detection runs per asset and raises an alert with the signals that drove it. The maintenance copilot takes the alert, retrieves the relevant manual pages, SOPs and the asset's fault history, and prepares a draft work order with likely parts.
The technician receives it on a tablet or phone, asks follow-up questions by voice or text, confirms the diagnosis and closes the work order in the CMMS. Their notes and the outcome flow back into the fault history, which makes the next retrieval better and gives the anomaly model its labels.
Nothing in this loop touches machine control. The retrieval layer follows our AI knowledge base pattern and the voice interaction our voice AI work.
Data flows out of the OT network, never commands in; the technician closes the loop and trains the next alert.
Track record
What we have built that applies
IoT products and companion apps
Connected devices with mobile and web companions, from precision agriculture tools to building and equipment monitoring, with the firmware-to-cloud plumbing that industrial data needs.
No. Safety and control systems stay deterministic and certified. AI reads from sensors, MES and the CMMS, and advises people: a flagged anomaly, a suggested cause, a draft work order. The boundary between the operational technology network and the AI services is a design input from day one.
We do not have years of labelled failure data. Can anomaly detection still work?
Yes. Unsupervised methods learn what normal looks like per machine and flag departures from it, which needs history but not labels. As technicians confirm or dismiss alerts, the labels accumulate and the model sharpens. We start with the assets that cost most when they stop.
Our manuals are PDFs from a dozen OEMs. Is that usable?
That is the typical starting point. We index manuals, SOPs, fault histories and shift notes so a technician can ask a question in plain language and get the relevant page and diagram, cited. See document Q&A.
Can it run on the plant network without internet?
Yes. Open-weight models and the retrieval stack can run on-premise or at the edge, with no outbound connection, where connectivity or policy requires it. See on-prem versus cloud LLMs.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.