SaaS Development Company in Munich

Industrial Manufacturing / IoT

A Munich-based industrial manufacturer (Mittelstand, €320 million revenue, producing precision components for automotive and aerospace — 4 factories in Bavaria with 2,800 CNC machines, grinding machines, and assembly stations) was losing €8 million annually to unplanned downtime. Equipment failures disrupted production schedules, created rush orders for replacement parts, and caused late deliveries that strained relationships with automotive OEMs (BMW, Audi) who required just-in-sequence delivery. The company had sensor data from some machines (vibration, temperature, power consumption) but no analytics — the data was collected but sat unused in historian databases. Problems: reactive maintenance (87% of maintenance was reactive — fixing machines after they broke), no failure prediction (maintenance teams relied on experience and scheduled intervals — a bearing might be replaced every 6 months regardless of actual condition), data silos (sensor data was trapped in individual machine controllers — no unified view across the 2,800 machines), and Betriebsrat concerns (the works council was concerned that any IoT analytics platform might be used to monitor individual worker productivity — a legitimate concern under German co-determination law). The company estimated that predictive maintenance could reduce unplanned downtime 40% and save €3.2 million annually.

We built a predictive maintenance SaaS with machine connectivity, AI failure prediction, maintenance workflow, and Betriebsrat-compliant analytics. (1) Machine connectivity: unifying data from 2,800 machines. Edge gateway deployment: IoT edge gateways installed at each factory — collecting data from machine controllers via OPC UA (for newer machines) and Modbus (for older machines). Edge processing: data pre-processed at the edge — filtering noise, aggregating high-frequency sensor data (some sensors produce 10,000 readings/second) into meaningful metrics, and detecting anomalies locally before sending to the cloud. The edge architecture ensures the system works even with intermittent internet connectivity. Unified machine registry: every machine catalogued with: machine type, manufacturer, installation date, maintenance history, and connected sensors. The registry provides a single view of the entire equipment fleet across 4 factories. (2) AI failure prediction: predicting breakdowns before they happen. Anomaly detection: ML models trained on each machine's normal operating pattern — detecting deviations in vibration, temperature, power consumption, and acoustic signatures that indicate developing problems. Failure classification: when an anomaly is detected, the model classifies the likely failure mode — bearing wear, spindle imbalance, coolant degradation, tool wear, or electrical fault — and estimates time to failure (hours, days, or weeks). The model was trained on 18 months of historical maintenance records correlated with sensor data. Confidence scoring: each prediction includes a confidence score — high confidence (>85%) triggers automatic work order creation, medium confidence (60-85%) creates an alert for maintenance engineer review, and low confidence (<60%) is logged for trend monitoring. Continuous learning: the model improves as maintenance engineers confirm or correct predictions — creating a feedback loop where every maintenance event makes the model more accurate. (3) Maintenance workflow: from prediction to resolution. Automatic work order: when a high-confidence prediction is generated, a work order is automatically created — specifying the machine, predicted failure mode, recommended action, parts needed, and estimated repair time. Spare parts integration: the work order checks spare parts inventory — if the needed parts are in stock, they're reserved. If not, an automatic purchase order is generated to the supplier with the appropriate urgency based on predicted time to failure. Maintenance scheduling: predicted failures are scheduled into the maintenance calendar — coordinated with production schedules to minimise impact. A bearing replacement predicted in 10 days can be scheduled during the next planned production break. Mobile for maintenance teams: technicians receive work orders on mobile devices — with machine location, failure prediction details, recommended procedures, and parts pickup location. Completion confirmation with photos and notes feeds back into the learning model. (4) Betriebsrat-compliant analytics: machine performance, not worker performance. Aggregate-only operator analytics: the platform analyses machine performance, not operator performance. All analytics are at the machine level — "Machine CNC-147 is producing parts 3% slower than its optimal rate" — not "Operator Schmidt is 3% below target." Betriebsrat-reviewed design: the platform's data model and analytics were reviewed by the Betriebsrat's technical advisor before deployment. The review confirmed: no individual operator identification in analytics dashboards, no correlation of machine performance with operator schedules (unless explicitly approved by the Betriebsrat for specific safety purposes), and audit logs designed for system security, not for operator activity monitoring. Betriebsrat agreement: a formal Betriebsvereinbarung (works agreement) was documented — specifying exactly what data is collected, how it's processed, who has access, and the purposes for which it may be used. This agreement is referenced in the platform's configuration and enforced technically.

Unplanned downtime: reduced 52% in the first 12 months (exceeding the 40% target). The AI model predicted 78% of unplanned failures with sufficient lead time (>48 hours) for scheduled repair. The remaining 22% of failures were detected by the anomaly detection but with less than 48 hours lead time — still enabling faster response than fully reactive maintenance. Cost savings: €4.8 million annually (50% above the €3.2M estimate). Breakdown: unplanned downtime reduction (€3.1M — calculated from production hours recovered at €380/hour average), spare parts optimisation (€800K — predictive ordering eliminated emergency shipping surcharges and enabled bulk purchasing), and maintenance labour efficiency (€900K — scheduled maintenance is 40% more labour-efficient than reactive maintenance). Prediction accuracy: 78% of predictions correct (machine failed within ±20% of the predicted timeframe). Accuracy improved from 62% in month 1 to 78% by month 12 as the model learned from maintenance engineer feedback. False positive rate: 8% (predictions that didn't result in failure) — acceptable to maintenance teams who confirmed that a precautionary inspection is far cheaper than an unplanned breakdown. Maintenance shift: reactive maintenance decreased from 87% to 38% of total maintenance activity. Predictive maintenance now accounts for 45%, with scheduled preventive maintenance at 17%. OEM delivery: on-time delivery to BMW and Audi improved from 94.2% to 98.7% — a significant improvement in the just-in-sequence delivery that automotive OEMs require. The improved reliability strengthened relationships with automotive customers. Betriebsrat satisfaction: the Betriebsrat chairman stated the platform was "a model for how industrial IoT can be implemented while respecting worker dignity and co-determination rights." The formal Betriebsvereinbarung is now referenced by other Bavarian manufacturers as a template. Platform expansion: based on the Munich success, the company deployed the platform at 2 additional factories (Regensburg and Augsburg) with minimal configuration — the machine-type-based models transferred well across facilities. Development investment: €420,000 over 16 weeks.