Run everything inside your network.
Fully air-gappedThe knowledge graph, agents and models run locally. Updates arrive as signed offline packages and no data leaves the site.
- Local model serving
- Read-only connectors
- Your SSO, roles and audit
The same research, applied to each domain’s evidence, rules and owners.
A worked example: in-line vision starts rejecting more parts. Agents separate real defects from false calls, tie them to the process and the lot, and a quality engineer decides. Every figure on this page downloads as SVG or PNG.
The questionWhy is the vision system rejecting more parts on line 2 since Monday, and are they real defects?
Pull every reject since Monday with its image, class, confidence and station.
Group by class and position on the part to find the signature.
Compare confidence and class mix with the reviewed history: real defects or false calls?
Line up the signature with station parameters, tool changes and lots.
Propose release, rework or scrap per lot, with example crops attached.
Most new rejects are real burrs on the inner edge after the station 4 tool change. 37 are glare false calls, queued for relabelling.
The quality engineer accepts, edits or rejects the proposed disposition for each lot.
Relabelled images join the training set only when reviewers agree. A new model version passes the eval before it runs on the line.
Defects cluster by class and position. A cluster that lines up with a station change is a process question, not a model question.
Frames, classifier output, station data and the defect catalogue are read in place and linked in one plant model.
Agent Hive
Agent HiveFrames and crops with station, timestamp and part ID.
Class, confidence and model version for every decision.
Station parameters, tool changes and genealogy from the MES.
A vision agent reviews the rejects while a process agent reads the station data. Their findings meet at the disposition, which a quality engineer approves.
Objects read off an image are grounded in the plant graph: the part, the station, the defect class and the limit that applies.
Boxes, polygons and keypoints on the frames reviewers flag.
Images where reviewers disagree go back for a second look.
The training set grows from reviewed images, with their history.
A new model runs on the line only after it passes the eval.
The eval set mixes real defects, known false calls and hard edge cases. The candidate must hold every case the line depends on.
| Question | m1 | m2 | m3 | m4 |
|---|---|---|---|---|
| Is this scratch inside the cosmetic limit?Classify | ✓ | ✓ | ✓ | ✓ |
| Glare on a flange: defect or false call?False call | ✗ | ✗ | ✓ | ✓ |
| Which station shows the burr cluster?Locate | ~ | ✓ | ✓ | ✓ |
| Did rejects rise after the tool change on Monday?Tie to process | ✗ | ~ | ✓ | ✓ |
| Which lots hold parts with the edge chip?Tie to lot | ✓ | ✗ | ✓ | ✓ |
| Pit under 0.2 mm on a sealing face: reject?Apply the limit | ~ | ~ | ✗ | ✓ |
| Show three example crops for the clusterShow the evidence | ✓ | ✓ | ✓ | ✓ |
Cases come from rejects a person has already judged.
Over-rejecting good parts fails the eval too.
Answers cite the acceptance criterion they applied.
The quality lead signs off a new model version.
Vision models and image stores often cannot leave the plant. Run fully local, or send only approved, redacted requests through a gateway.
The knowledge graph, agents and models run locally. Updates arrive as signed offline packages and no data leaves the site.
A policy gateway controls what is sent, which model receives it and what comes back into the workflow.
Point at a part 4 sites · 1 control plane · data stays in its region