The Challenge
Kapture IO was a SaaS platform for quality control in industrial manufacturing. Its clients ran high-volume production lines where undetected defects could cost hundreds of thousands of euros. One client manufacturing automotive bumpers had a 10% scrap rate on 10,000 units/day (roughly €7.5M/year in waste). A single undetected issue running for 48 hours meant 20,000 defective parts and €600,000 in losses from one incident alone.
The platform collected quality data from factory floors but couldn’t act on it fast enough. Detection took days. We also identified data integrity issues in the quality process itself: impossible trip paradoxes (a sample inspected at two locations too far apart given the time between records) and temporal inconsistencies suggesting errors or manipulation.
As Product Manager I partnered with Eurecat (a technology center with applied AI capabilities) to bring ML-powered anomaly detection into the platform:
- Eurecat owned the ML model (training, validation, performance)
- Our team owned the product integration, UX and infrastructure
- My job was bridging both teams and making sure the feature worked for quality inspectors on a factory floor, not data scientists
The Solution
- Backend pipeline processing thousands of quality samples daily, classifying anomalous patterns in real time
- Validation interface for edge cases where human review improved model accuracy through feedback loops
- Monitoring dashboard for factory managers showing anomaly rates, trends and drill-down into specific incidents
The Results
- Quality anomalies flagged within hours instead of 2-3 days
- Up to €600,000 in scrap prevention per incident for high-volume lines
- Systematic identification of data integrity issues previously invisible in manual reviews
- First ML-powered anomaly detection in the industrial quality control SaaS space