Skip to content
Blueprint · Manufacturing

Visual inspection, explained. From camera frame to disposition.

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.

Manufacturing
The question

Why is the vision system rejecting more parts on line 2 since Monday, and are they real defects?

Inputs

What the agents read.

Inspection images
Frames and crops from in-line cameras, by station
Classifier results
Defect class, confidence and model version per part
Process data
Station parameters around each reject
Defect catalogue
Reference images, limits and acceptance criteria
Labelled history
Reviewed rejects and confirmed false calls
Run

What happens, step by step. Each step is logged with its sources.

  1. 01

    Gather the rejects

    Pull every reject since Monday with its image, class, confidence and station.

  2. 02

    Cluster the defects

    Group by class and position on the part to find the signature.

  3. 03

    Check the model

    Compare confidence and class mix with the reviewed history: real defects or false calls?

  4. 04

    Tie it to the process

    Line up the signature with station parameters, tool changes and lots.

  5. 05

    Draft the disposition

    Propose release, rework or scrap per lot, with example crops attached.

Agent Hive · Run 3381
Why is the vision system rejecting more parts on line 2 since Monday?
  1. Pulled1,284 rejects with crops2.2s
  2. Clusteredby class and position3.4s
  3. Comparedwith 9,800 reviewed images1.6s
  4. Checkedstation 4 tool change, Monday 06:100.7s
  5. Drafteddisposition for 6 lots0.5s
Answer4 sources
+3.1 pts4.6%
Reject rate, line 2, since Monday

Most new rejects are real burrs on the inner edge after the station 4 tool change. 37 are glare false calls, queued for relabelling.

vision.resultsimage storestation_4 logdefect catalogue
Ask a follow-up…
Approvals

Where people decide.

  • Needs you

    Disposition

    The quality engineer accepts, edits or rejects the proposed disposition for each lot.

  • Needs you

    Labels and model changes

    Relabelled images join the training set only when reviewers agree. A new model version passes the eval before it runs on the line.

Outputs

What you get back.

Reject review
Defects by class and position, with example crops
Lot disposition
Release, rework or scrap, with the evidence
Labelling queue
Images where reviewers or the model disagreed
Run trace
Every image, query and model version used
Evidence

Find the signature. Then follow it upstream.

Defects cluster by class and position. A cluster that lines up with a station change is a process question, not a model question.

Fig. 1 · Defect mapIllustrative data
Rejects classified and placed on the part. The cluster in violet lines up with the station 4 tool change.
Fig. 2 · Question with image evidenceIllustrative data
A reviewer asks a question about one camera frame; the answer marks the image region it used as evidence.
Data sources

Images are evidence too. They sit beside the process data.

Frames, classifier output, station data and the defect catalogue are read in place and linked in one plant model.

Fig. 3 · Connected sources
Sources stay in place; Agent Hive reads them through read-only connectors and builds one shared model. Logos are trademarks of their respective owners.
  • Image store

    Frames and crops with station, timestamp and part ID.

  • Inspection results

    Class, confidence and model version for every decision.

  • Process and lot data

    Station parameters, tool changes and genealogy from the MES.

Agent workflow

Look, compare, then decide. Disposition waits for a person.

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.

Fig. 4 · Workflow runIllustrative data
The run as the platform draws it: tool steps, specialist agents, joins, branches and the approval it is waiting on. Drag to pan; Ctrl or ⌘ + wheel to zoom.
Knowledge model

What the image shows, linked to what the plant knows.

Objects read off an image are grounded in the plant graph: the part, the station, the defect class and the limit that applies.

Fig. 5 · GroundingIllustrative data
Objects read off a frame, linked to entities in the plant graph: part, station, defect class and the limit that applies.
Labelling loop

Labels improve the model. Reviewers keep them honest.

  • Label

    Mark what the model missed

    Boxes, polygons and keypoints on the frames reviewers flag.

  • Agree

    Check agreement first

    Images where reviewers disagree go back for a second look.

  • Train

    Add agreed labels only

    The training set grows from reviewed images, with their history.

  • Gate

    Grade before release

    A new model runs on the line only after it passes the eval.

Fig. 6 · LabellingIllustrative data
Boxes, polygons and keypoints on the frames reviewers flagged.
Fig. 7 · Reviewer agreementIllustrative data
Labels enter the training set only where reviewers agree. Disagreements go back for a second look.
Evaluation

Every model and prompt is graded. On images reviewers already checked.

The eval set mixes real defects, known false calls and hard edge cases. The candidate must hold every case the line depends on.

Fig. 8 · Question set by versionIllustrative data
Evaluation results by question and version
Questionm1m2m3m4
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✓✓✓✓
m4: 7 of 7 pass · release needs all 7
Each row is a question agreed with the team; each column a version of the workflow. The candidate ships only if it holds the release bar.
  • 01

    Reviewed images only

    Cases come from rejects a person has already judged.

  • 02

    False calls count

    Over-rejecting good parts fails the eval too.

  • 03

    Limits from the catalogue

    Answers cite the acceptance criterion they applied.

  • 04

    Owner approves release

    The quality lead signs off a new model version.

Outcomes

Fewer false calls. Faster, defensible dispositions.

False rejects
−38%
after reviewed relabelling
To a lot disposition
20 min
with example crops attached
Reviewer agreement
0.81
kappa on the labelling queue
Eval pass, m4
7 of 7
release bar for the line
Illustrative data
Deployment

Images stay on site. Models can too.

Vision models and image stores often cannot leave the plant. Run fully local, or send only approved, redacted requests through a gateway.

Run everything inside your network.

Fully air-gapped

The 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

Keep the data local, use approved models.

Hybrid model access

A policy gateway controls what is sent, which model receives it and what comes back into the workflow.

  • Model choice per agent role
  • One controlled egress path
  • Graph and sources stay local
Fig. 9 · Hybrid model access
Multi-cloud and hybridThe same platform runs in several clouds and regions and on-prem. A control plane sends versions, policies and releases and receives health; each site keeps its data, context graph, models and audit log in its own region, and fails over only inside it. Point at a site.Control planeversions · policies · releaseshealth ↑ · no customer dataCLOUD A · EUREGION EURUNTIMEzone 1activezone 2standbyfailoverin-regionData · context graphat rest in-regionModelsapproved endpointsAudit logkept in-regionDATA RESIDES IN EUCLOUD B · INREGION INRUNTIMEzone 1activezone 2standbyfailoverin-regionData · context graphat rest in-regionModelsapproved endpointsAudit logkept in-regionDATA RESIDES IN INDIACLOUD C · USREGION USRUNTIMEzone 1activezone 2standbyfailoverin-regionData · context graphat rest in-regionModelsapproved endpointsAudit logkept in-regionDATA RESIDES IN USON-PREM · SITEON SITERUNTIMEzone 1activezone 2standbyfailoverin-regionData · context graphat rest in-regionModelslocal or via gatewayAudit logkept in-regionDATA STAYS ON SITENO CROSS-REGION DATAneutral cloud outlines · no vendor marks · regions are examplesdashed: configuration and releases down, health up

Point at a part 4 sites · 1 control plane · data stays in its region

Images and the plant graph stay on site. Only approved, redacted requests leave through one gateway.

Run this blueprint on your data. We set up the first one with you.