An inspection cell that had to reach a verdict, drive a reject and attach the result to a batch record — all inside a cycle time the line would not extend. The model was the smallest part of the work.
The defect class was real, costly and inconsistently caught by manual inspection. It was also, in the first images supplied, almost invisible — which is the usual starting position and the reason the engagement began with an imaging trial rather than a model.
The line would not slow down. The station cycle was fixed, mechanical motion consumed most of it, and whatever remained was the entire budget for capture, transfer, inference and decision. That number constrained camera resolution, model architecture and inference hardware simultaneously, before any of them were chosen.
There was also a PLC that had run the line for years and was not being replaced, an MES that expected results attached to batch records, and an operations team who had seen a previous vision pilot fail.
Stated as the customer stated them, before any of them had an answer. A challenge that is only described after it was solved is a description of the solution.
Under the diffuse illumination already installed it produced no usable contrast at any resolution. No model recovers information that was never captured.
Whatever remained after mechanical motion was the entire budget for capture, transfer, inference and decision — and it constrained every hardware choice at once.
A reject station three metres downstream means encoder-based tracking through line stops and restarts. A correct verdict applied to the wrong part is worse than no inspection.
The operations team had seen a vision pilot fail before. Any system would have to prove itself against existing inspection rather than replace it on assertion.
The illumination block is marked because changing its geometry was the single highest-value decision in the programme — everything downstream depended on the defect being visible at all. Part tracking is marked because a verdict that reaches the wrong part is worse than no inspection.
The specific scope, rather than a capability list. Where a stage was shared with the customer’s team, it is described as shared.
Rarely the subsystem that sounds difficult. Written out because a reader facing the same programme gets more from this than from a summary of what went well.
Under the illumination already installed the defect produced no usable contrast. No model, resolution or architecture would have recovered it. Changing the geometry to low-angle was the single highest-value engineering decision made.
Average inference time fitted the budget comfortably. The ninety-ninth percentile did not, until the input pipeline was restructured — and the ninety-ninth percentile is the one that misses a trigger and drops a part.
A shift register that loses synchronisation during a stop and restart sends a verdict to the wrong part. Handling that correctly took more effort than the model did.
Several defect modes had too few examples for supervised training and were handled by anomaly detection against good parts, rather than by pretending the class imbalance was not there.
The cell ran in parallel with existing inspection, visible to operators, long enough to produce a comparison they believed. Given a previous pilot had failed, scepticism was rational and had to be answered with evidence.
Verdict, reject actuation and MES record all completed within the station cycle, with worst-case latency measured on the actual hardware.
Documentation, monitoring dashboards and a retraining pipeline the site can run without Faststream.
Input statistics, output distribution and reject rate tracked per station, because a fleet average hides a single failing cell.
Imaging first, algorithm second, integration third. Reversing that order is the most common and most expensive mistake in industrial vision.
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, register maps, state machines and confidential deliverables are not disclosed. Where a figure would identify a customer or a design, it is omitted rather than approximated. More detail is available under a non-disclosure agreement, within the limits each customer has agreed.
Every item links to its own page, with characteristics, applications and the maturity status stated honestly for that item.
The full solution, architecture to deployment.
PLATFORMThe machine vision platform.
PRODUCTThe hardware inference ran on.
AIImaging chain and illumination geometry.
AIDrift monitoring and retraining.
KNOWLEDGEHow the handshake works.
Because a defect that produces no contrast under the installed illumination contains no information for a model to find. Here the defect was undetectable under the original diffuse lighting and obvious under low-angle illumination, which determined everything downstream.
Because a trigger missed at the ninety-ninth percentile drops a part. Average latency fitting the budget is not sufficient; the tail is what the cycle time has to accommodate.
Drift monitoring on input statistics, output distribution and reject rate per station, a review queue for low-confidence cases that generates labelled data, and a retraining pipeline the site runs itself.
By anomaly detection trained on good parts rather than supervised classification, which needs examples the process had not produced.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.