Industry · Inspection

Computer vision & industrial inspection

Automated inspection that does not just find a defect. It localizes it, explains it and routes the part, at every step of the line, learning only from good parts.

01The problem

Take a semiconductor or PCB line. Defects can appear at any step, and a bad unit only gets more expensive the further it travels. Yet real defects are rare, so there is almost nothing to train a classifier on.

Inspection at every step. Take a semiconductor fab line as an example. An unsupervised inspector sits at every station and catches defects the instant they appear; here the etch step flagged one. The same inspector runs on PCB, MEMS, biomedical and assembly lines.

02How we solve it

Learn what “normal” looks like, then flag anything that deviates.

Each station runs an inspector trained only on good parts, so it never needs an example of a defect.

Teacher-student anomaly detection. Trained on good parts only. At inference, the student’s disagreement with the frozen teacher becomes a per-pixel anomaly map.

Teacher-student anomaly detection

A frozen teacher network and a student trained only on good parts. Where the student disagrees with the teacher, a defect is flagged, per pixel.

Memory-bank methods (PatchCore, PaDiM)

Store the features of normal patches and score each new one by its distance from normal. State-of-the-art accuracy on MVTec benchmarks.

EfficientAD

A lightweight variant that scores an image in milliseconds, so every unit is inspected inline at full line speed.

Reverse distillation (RD4AD)

A one-class encoder-decoder that reconstructs normal features, giving cleaner localization and fewer false positives.

03What it produces

Localized, explainable, and benchmark-anchored.

Per-pixel localization. The model does not just say “defect”. It shows exactly where, as a heatmap overlaid on the part.

Benchmark-anchored accuracy. A low-false-positive operating point keeps the line moving while catching real defects, validated against MVTec-style benchmarks.

Every unit is inspected at every step with pixel-level localization. Defects are caught the instant they appear, false alarms are filtered, and every decision traces back to evidence, with no labelled defects ever needed.

04The agentic loop

Detection becomes a decision.

Agentriq places an unsupervised inspector at every station and wraps each detection in an agent loop that re-checks it, explains it, and routes the part.

From detection to decision. Each flag is re-checked, explained against process logs, and routed to pass, rework or scrap, then written back to the line.

01Score every unit inline; the instant a patch exceeds threshold, raise a flag with a localized heatmap.
02A verification agent re-checks the flag with a second method to suppress false alarms.
03A root-cause agent and Agentic RAG correlate the defect signature with process and equipment logs.
04A judge agent decides pass, rework or scrap.
05The outcome and its evidence are written back to the line.

05Tooling

Simulation, solvers & frameworks.

Get started

From engineering problems to autonomous solutions.

Talk to our engineers john.ferderigos@agentriq-sa.com