Priya builds vision and process-monitoring systems that can explain themselves to a quality engineer. She is allergic to demos that only work on the training set.
Startups like her because she will say no early. Universities like her because the papers still come.
Her methodological insistence is on calibration. A model that reports ninety per cent confidence should be right about nine times in ten at that confidence, and most industrial models are not. Once calibration is established, a sensible hand-off rule to a human inspector follows almost automatically, and the system becomes something a factory can be accountable for.
She is equally insistent about test data. Her projects budget explicitly for labelling and for a held-out set gathered on a different week, shift and material lot from the training data. This is unglamorous, it consumes perhaps a third of the effort, and it is the difference between a deployment and a demonstration.
Selected projects
Named work, with a sentence of context each.
Weld inspection with calibrated hand-off
Deployed a vision system that refers low-confidence welds to a human inspector, and reduced the teardown rate without ever claiming to have removed the inspector.
Independent model validation
Validated a supplier's defect-detection model on data it had not seen, and found that performance on a second production site was materially worse than the tender had implied.
Drift monitoring after a material change
Built a drift detection layer that flagged degradation within days of a supplier substitution, in a plant where the previous system had quietly deteriorated for months.
Working with Venkat
She will want to walk the line before she looks at the data, and to speak to the quality engineer rather than the project sponsor. What counts as a defect is a human judgement, and it is usually less consistent than anyone expects.
Her projects begin with an agreed success metric and an agreed cost for each kind of error. If those cannot be settled, she regards the project as not yet ready, and says so.
Looking for
Messy production data and a plant manager who will walk the line with you.
Probably not for
Proofs of concept intended for a slide rather than a line. She will decline politely and early.
Credits and contributions
- Author of an internal guidance note on calibration adopted by two manufacturing partners
- Reviewer for machine learning and manufacturing conferences
- Developer of an open benchmark for industrial defect detection under distribution shift