04 — Manufacturing — 2024

Sable

for Vitro Kilnworks

Vitro fires architectural tile in Sacramento. Hairline glaze cracks were being caught at packing rather than at the kiln exit, which meant scrapping a full pallet instead of one piece. We put a model on the line.

Plate 04 — generated diagram, sable telemetry

96.8%
Recall on hairline cracks: 41ms
Inference, on-device: $214k
Annual scrap avoided

Small model, hard problem

There was no budget for a GPU on the line and no appetite for sending kiln footage to a cloud. The whole system had to run on a compute module bolted to a conveyor housing, in an environment that reaches 44°C.

We labeled 6,400 tiles with Vitro’s own quality lead — a genuinely tedious three weeks that produced a far better dataset than any synthetic augmentation strategy we tried. The final network is 4.1 megabytes.

The operator is the point

A model that flags a defect is worth nothing if the person at the line does not believe it. Sable shows the cropped region it reacted to, at full resolution, next to the two most similar confirmed defects from the archive. Operators overrule it about four percent of the time, and every overrule is training data.

“Accuracy got us to a demo. Showing our work got us into production.”

— Devendra Raman, applied ML

Next in the index — 05

Pressroom

An editing environment built around the copy desk, not the CMS vendor.