ML-Powered Failure Detection for a Diagnostics Manufacturer

The situation

A diagnostics manufacturer produces cartridges used in medical testing equipment. Quality control is governed by regulatory requirements: if too many cartridges in a sampled lot fail testing, the entire lot must be discarded. Failures stem from defects in ultrasonic welding and film sealing processes, both of which produce rich sensor data and infrared imagery during production. But that data was not being used for early failure detection. Engineers discovered defective cartridges only during lot-level quality control, after the cartridges had already been incorporated into lots subject to regulatory sampling.

The company’s data science team had developed prototype ML models that showed promise for predicting weld and seal failures from production telemetry. What they lacked was the infrastructure to train those models on integrated datasets, deploy them to production, and serve predictions back to the manufacturing line at the latencies required for in-line quarantine decisions.

What we built

CorrDyn conducted an assessment that included 8+ stakeholder interviews, architecture and code review, infrastructure alternatives research, and a working proof of concept. The output was a data infrastructure roadmap and the initial implementation of two parallel systems.

Training data pipeline. Manufacturing machines send sensor data through PLC and OPC connections to Ignition, the plant’s supervisory control system. CorrDyn added a streaming path from Ignition to AWS Kinesis Data Firehose, which lands data in S3 and loads it into Snowflake via Snowpipe. This gives the data science team a warehouse with integrated relational data (weld events, sensor time series, seal test results) and references to infrared camera imagery stored in S3. Data quality checks validate PLC identifiers, serial number uniqueness, image counts, and label integrity before data enters the training pipeline.

Model serving architecture. We designed a containerized model serving system with three layers. Manufacturing machines trigger prediction requests through Ignition. A business logic layer (middleware) receives the request, preprocesses the data, selects the appropriate model based on the event type, and routes the request to the correct model inference endpoint. Each model runs as a containerized FastAPI application deployed on AWS ECS. The middleware returns predictions to Ignition, which can trigger automated cartridge quarantine on the production line.

The architecture supports multiple models per machine (weld failure, seal failure) and is designed to scale from one production line to dozens across multiple facilities. A CI/CD pipeline through GitLab handles model versioning, testing, and deployment.

What changed

The manufacturer gained the ability to flag potentially defective cartridges during production rather than discovering them during lot-level quality control. Quarantining suspect cartridges before lot assembly reduces the probability of regulatory lot failures, saving both materials and production time. The data science team can now develop models against integrated datasets in Snowflake and deploy them through a standardized process rather than ad hoc handoffs. The infrastructure was designed for expansion: bringing a new machine or model type online follows the same pipeline pattern without requiring new architecture.