COMPUTER VISION · SPEQ SYNTHESIS
ML Automated Visual Inspection
A machine-learning computer-vision system that inspects units on the manufacturing line, classifying defects such as cracks, particulates, fill anomalies, or cosmetic faults faster and more consistently than manual inspection alone.
Because its output feeds accept and reject decisions on product that may reach patients, the model’s error profile — especially the defects it misses — is a direct product-quality concern, not a convenience metric.
What a system class is not
An AI system class describes a shape of system, not a product and not an approval pathway. SPEQ does not qualify, validate, or endorse any implementation, and no regulator recognises these classes as a category.
Weighted score 25/32 (decision consequence and model influence weighted most heavily) places this in the high tier.
Computed deterministically from four Context-of-Use factors — decision consequence, model influence, data sensitivity, and change dynamics. A transparent scoping aid, not a validated risk-assessment system.
SCOPE THE ASSURANCE STRATEGY → CSA WORKBENCHDeployed on the production line to classify units as pass or reject, influencing batch disposition, defect rates, and ultimately whether specific product is eligible for release.
Its classifications directly affect which units are rejected or released, so an undetected miss or systematic misclassification can release defective product or discard conforming units against the batch record.
human in the loop
Because misses affect product that reaches patients, qualified human review of rejects and defined escalation for uncertain cases must remain in the loop, and disposition accountability stays with a person.
- A false negative that passes a genuine defect is the most consequential failure, because it can release non-conforming product while the automated result reads as a clean pass.
- The model can drift as lighting, cameras, materials, or a new defect type shift the production distribution away from the data the model was trained and qualified on.
- Performance can degrade unevenly across defect classes, so overall accuracy stays high while a rare but critical defect type becomes systematically under-detected.
- Automation bias reduces the vigilance of human inspectors who now only see flagged units, weakening the backstop that is meant to catch model errors.
- Opaque model behavior makes a specific accept or reject call hard to justify to an investigator, complicating any defense of the disposition decision.
- Qualify the model against a representative, labeled challenge set spanning every defect class and matrix, reporting sensitivity per defect type rather than a single aggregate accuracy figure.
- Define and monitor the operating point deliberately, treating false negatives on critical defects as the governing risk when the accept and reject threshold is set.
- Monitor input and prediction distributions in production to detect drift from lighting, materials, or new defect types, and re-qualify when the process changes materially.
- Establish a human review and escalation path for uncertain or borderline classifications so ambiguous units are adjudicated by a qualified inspector rather than auto-disposed.
- Maintain data integrity and attributable records linking each inspected unit, its model result, and its final disposition so the batch record remains defensible.
Standards SPEQ maps to this class
AI-governance frameworks
The AI-specific shelf that defines “quality AI” — see Good AI Practice.
SPEQ synthesis — applied AI-assurance judgment to help you scope your own Context-of-Use assessment and validation. Not regulatory guidance, not an AI classification service, and not a substitute for your documented risk assessment.