Grading a crop that changes with the season

A food plant changes with the season in a way few other floors do. The crop that comes through the door in spring is not the crop that comes in autumn, the pack changes with the customer, and sometimes a whole line is reconfigured between them. The grading and inspection cameras on those lines came from different equipment vendors, and each carries a model trained elsewhere, on someone else's produce. The plant's own people know what good looks like this week; the cameras are working from a memory of a different plant. A Czar Grid in the plant adapts the models to this plant's crop on this plant's captures, and serves every line from the same place.

What is on the line

A processing floor is a sequence of vendors. The grader came from one, the packer from another, the foreign-object detection from a third, and the palletizer at the end from a fourth. Beyond the line are the cold rooms, the dock, and increasingly the mobile robots that move product between them. Each camera on the floor is its own island: its own model, its own update schedule, its own account with its own vendor, and no way to tell the next camera what it has learned about this week's crop.

The pool does not ask the plant to replace any of it. Each vendor's software calls the Grid over the plant's own network for the models it cannot run on its own hardware, within the time budget the line has. Safety, stops, and motion limits stay on the robot as the vendor built them; the pool answers questions, it does not drive the machine.

Why a model trained elsewhere and a GPU on every machine both fail

A model trained on another plant's produce grades this plant's produce badly, and the usual remedy is worse than the problem: send this plant's imagery to the vendor so the vendor can improve a model it will then serve to everyone, including the plant down the road. Line imagery is the plant's data and its best training material, and under that arrangement it leaves the building to improve someone else's camera. Each vendor's cloud is a separate uplink, and each vendor's model learns the season on its own schedule, if at all.

A GPU on every machine fails on the floor itself. A wash-down line is hosed, a cold room is cold, and neither is a place to keep hardware running hot and alive. Capacity sized for each camera's busiest moment sits idle between them and cannot be shared with the next line.

Grading and packing, one pool

On the Grid, Inference serves the grading, packaging, and foreign-object models from the plant's own rack to whichever line asks, from any equipment vendor, over the plant's network. The frames a camera sends to be read are read in the plant and stay there. Memory keeps what the pool has learned about the current crop across lines, so the grader at the front and the pack check at the back are working from the same picture of what came in this morning.

Local Bridge permits only the routes the plant declares, typically the plant's execution or quality system and the equipment vendors' line software, and denies the rest by default. On an air-gapped plant there is no route out, and updates arrive as signed bundles carried into the room.

A line that learns the season's crop

The season is where the pool earns its place. Adaptation tunes the models on the plant's own captures as the crop changes, on the pool, overnight, so the first week of a new variety is learned from this plant's lights, this plant's belts, and this plant's idea of a reject. Evaluation keeps a held-out set of the plant's own images and checks each adapted model against it before Inference serves it, and the change is written to the record with the evaluation beside it. The plant's quality lead can see what the model was asked to call good last week and this week, and why.

Putting an adapted model into service on a line is a task the plant can bound. In Agency Studio it is authored once, approved by a named role, and recorded, so the morning a new grade threshold goes live is on the record with the person who approved it.

What stays in the plant

Imagery, adapted models, task definitions, and the record sit on the plant's own hardware, under the plant's own access rules. Watchman verifies the integrity of what the nodes are running. The plant's captures improve the plant's models and nobody else's; a customer's pack specification and a season's reject rate are the plant's business, held on the plant's hardware, read by the plant and by anyone the plant invites.

Placing and buying the pool

The nodes go in the plant's own rack, in the conditioned room that already holds the network equipment and the plant systems, on the warm and dry side. The plant network carries inference to the lines, the cold rooms, and the dock. The subscription is for the configured Grid: hardware, platform license, and sustainment in one line, fixed for the term, and it does not move when a line is added or an equipment vendor changes. Where an equipment vendor would rather ship the pool with its line under its own name, it does so as Czar Chassis, which is the same Grid reaching the plant by another route.

A plant that runs its vision this way stops relearning its own crop every season from scratch on every camera. The crop still changes. What changes with it is the plant's own model, on the plant's own captures, in a rack the plant can walk to.