Manufacturing & Industry

Whether a line runs profitably, whether an order will ship on time and where the scrap comes from is not written down in any single system. The machine delivers timestamps and sensor values, the MES delivers confirmations and piece counts, the ERP delivers orders and costs. In between sits a lot of manual work in spreadsheets.

We build data platforms and business intelligence that give the plant and management the same numbers. AI and machine learning add to that where their value can be tested on the company’s own data.

Machine, order, time period

The sticking point between shop floor and ERP is the shared key. Machine data knows no production order, and business systems know no sensor IDs. Only a maintained mapping of machine, order and time period turns the two worlds into one analysis. Building it usually takes longer than the technical connection over OPC UA, MQTT or the historian.

Older machines are mostly not an obstacle. Their controllers can often be connected through gateways or retrofitted sensors, and since the data is only read, we do not touch the control logic. We start with the machines needed for the first question and expand from there. We plan the connection with plant IT, because the security of the production network comes first.

Machine data is mostly not personal. As soon as it can be linked to individuals, for example through shift logins, the GDPR and codetermination rights apply.

OEE is not the same everywhere

Two plants report an OEE and mean different things. One counts changeovers as planned downtime, the other does not, and each uses a different reference cycle time. For scrap, it makes a big difference whether rework is counted. We agree on these definitions with production and finance together and store them once in the data model.

Data capture at the line matters just as much. If operators always pick “Other” as the reason for downtime, there will be no root cause analysis later, however good the dashboard is. That can only be solved together with production, for instance with a shorter selection list.

Once the definitions are settled, metrics can be broken down to the order, the batch or the machine. Combined with cost and calculation data, you can see what an order actually cost. With small lot sizes, it becomes especially visible how much setup time contributes.

Warehouse or lakehouse

For order, quality and cost data, a data warehouse is usually enough. A lakehouse pays off when large volumes of sensor time series, images or logs arrive that fit tables poorly, and when they arrive as a stream. The extra complexity is only worth taking on if the planned analyses need it. Which machines and questions come first, we settle in a data strategy.

Predictive maintenance: thresholds first, then a model

A forecasting model learns from documented failures. When failures are rare or their causes are missing from the maintenance log, it is hard to train. Condition monitoring with thresholds is then the better start. Whether a model detects anomalies early enough and with few false alarms, we test beforehand on the existing data.

A successful pilot is not yet a solution for every machine either. A model built for one machine rarely carries over unchanged to another.

Visual inspection with cameras

Image-based inspection works for clearly defined tasks with stable capture conditions, for example consistent parts and constant lighting. What matters are training images that include the rare defects. In operation, quality staff spot-check the results, and the model is re-tested for new products or changed lighting. If the AI is part of a safety function, the EU AI Act may apply. We clarify that case by case with legal counsel.

Data Act and sustainability reporting

The EU Data Act has applied since 12 September 2025. Users of connected products have access to the data their product generates. For new connected products placed on the market from 12 September 2026, manufacturers have to build this access in at the design stage, where relevant and technically feasible. Operators of machines can request access to the data of their equipment. Which products and data are affected is a question for legal counsel, and we help implement the technical side of providing the data.

On sustainability reporting, the EU has limited the CSRD obligation to large companies through the Omnibus I directive, to our knowledge to companies with more than 1,000 employees and more than 450 million euros in net turnover. The reporting standards (ESRS) were simplified too. Energy and emissions data per unit can be a basis for such reports, voluntary ones included. Whether and to what extent a company has to report is for legal counsel to say.