Plumbline builds process monitoring and inspection models for manufacturers, and has built a business on a fairly narrow proposition: their models report uncertainty honestly and hand the difficult cases to a person.
This sounds modest and is commercially significant. Their customers have generally bought an industrial AI system before, watched it degrade quietly after a material change, and lost confidence in the category. A system that flags its own drift and refers low-confidence cases is something a quality manager can be accountable for.
The company's least popular practice is insisting that a project budget for labelling and for a test set collected on a different week from the training data. This typically consumes a third of the effort and is, in their experience, the whole difference between a deployment and a demonstration.
They work in manufacturing and, increasingly, in infrastructure inspection, where the same principle applies and the consequences of an over-confident model are rather more serious.
Capabilities
- Process monitoring and defect detection
- Model calibration and uncertainty reporting
- Drift detection and post-deployment monitoring
- Independent validation of third-party models
- Edge inference on industrial hardware
- Labelling programmes and test set design
Ways in
The routes that exist, and what each is suited to.
Paid pilots with a defined success metric
A pilot with an agreed metric, an agreed cost for each kind of error, and a decision at the end. They decline unpaid pilots on principle.
Published work packages
Research questions put to the network, generally on calibration, drift or benchmark data they cannot generate alone.
Independent validation engagements
Assessing somebody else's model on data it has not seen. A short, unwelcome and frequently decisive piece of work.
How to approach them
They ask to walk the line and to speak to the quality engineer rather than the sponsor. What counts as a defect is a human judgement, and it is usually less consistent than the specification implies.
They will not begin without an agreed success metric and an agreed cost of error. Their view is that a project without those is not a project but an experiment with a budget.
Known for
- A weld inspection deployment with a calibrated hand-off to human inspectors, now in routine use
- A drift monitoring layer that detected degradation within days of a supplier substitution
- Independent validation work that has, more than once, ended a procurement