Vision inspection follow-through — secondary direction
What happens after the camera flags a defect
Sample review, defect re-checking, work orders, rectification, reporting and a management back office around a vision inspection system.
Operating context
Detection is one step in a longer chain
A vision system produces a signal. Somebody still has to look at borderline cases, decide what happens to the part, record the decision and answer for it later.
That second half — review, disposition, rectification, traceability and reporting — is where our workflow experience applies.
It is also where a plant is most exposed: if the record of who accepted what is thin, the value of the detection is reduced.
Objects the work is attached to
- Production line, station and batch
- Part or unit under inspection
- Detection event and image
- Human re-check decision
- Work order and rectification
- Quality report and audit trail
People involved
- Operator
- Quality inspector / re-checker
- Quality manager
- Maintenance for line-side rectification
Common operational problems
Where vision projects lose value
Borderline cases have no owner
Uncertain results need a human decision, and that decision needs to be recorded against the part.
Disposition is not traceable
Scrap, rework and accept decisions are made but not reliably linked to the image and the reviewer.
Rectification is disconnected
A recurring defect signals a process or tooling problem, but the link from detection to maintenance action is manual.
Reporting is rebuilt by hand
Quality reporting is assembled from exports rather than generated from the record.
Proposed workflow
Re-check, disposition, rectification, report
We work on the workflow around the algorithm, not on claiming the algorithm.
- 01Task
- 02Capture
- 03Verify
- 04Approve
- 05Sign
- 06Report
- 07Deliver
- 08Follow Through
- Task — a detection event or a sampling task is raised against a part, station and batch.
- Capture — the image, detection metadata and any additional operator observation are recorded.
- Verify — borderline and repeated cases are routed to a human re-checker.
- Approve — an authorised inspector records the disposition and the reason.
- Sign — the quality manager confirms where the disposition requires it.
- Report — quality reporting is generated from the recorded decisions.
- Deliver — the record is available for internal and customer quality review.
- Follow Through — recurring defects become maintenance, tooling or process tasks.
What the system can capture
Field evidence, structured at the point of work
- Detection event, image and station
- Part, batch and production context
- Human re-check decision and reason
- Disposition: accept, rework, scrap, quarantine
- Rectification work order and result
- Recurring defect patterns by station and shift
- Sample review sets for model tuning by the algorithm owner
- Audit trail of who decided what, and when
Capture is designed to support the engineer, not to police them: it should be faster than writing the same information twice.
Review and approval roles
Who decides what
Operator
Records what was observed at the station.
Quality inspector
Re-checks flagged and borderline cases and records the disposition.
Quality manager
Owns the final quality decision and what is reported to the customer.
Decision boundary. Final quality decisions are made by the customer’s authorised personnel. The detection algorithm may be supplied by a third party selected and authorised separately by the customer.
Reports and follow-through
A quality record that can be audited
The output is a defensible record: image, decision, reviewer, reason, action.
- Defect and disposition report by batch
- Recurring defect summary by station
- Rectification work order history
- Audit trail for internal and customer review
What happens next
- Recurring defects become maintenance or tooling tasks
- Reviewed sample sets go back to the algorithm owner for tuning
- Quality trends feed process improvement
- Customer quality queries can be answered from the record
Evidence status
What is proven and what is not
Workflow, review and back-office design — Prototype
The re-check, disposition, rectification and reporting workflow is designed and partly built. It has not completed acceptance in a production plant with a live line.
Vision algorithm supply — third party, customer-selected
Defect detection algorithms may be supplied by a third party that the customer selects and authorises separately. JJ AI TECH does not present those algorithms as its own.
Shenshuyin (深树阴) relationship — boundary statement
Shenshuyin is a company our founder has invested in. It is not an internal JJ AI TECH team or department. JJ AI TECH does not own its algorithms, customers, patents or results, and does not publish them as its own track record.
Named plant deployments — not claimed
We do not claim deployment at any named manufacturer, semiconductor or wafer customer.
Limitations
What this workflow does not do
- This is a secondary direction. Our primary focus is field service, maintenance, facilities and asset-intensive operations.
- We do not claim 100% accuracy, and we do not publish detection rates.
- We do not own or resell third-party vision algorithms without the owner’s written authorisation.
- Final quality, safety and customer-acceptance decisions remain with the customer’s authorised personnel.
- No production deployment at any named customer is claimed on this site.
If you already have detection, start with the review loop
The most useful conversation is about who re-checks a borderline result, what they are allowed to decide, and what record has to survive the audit.