AUTO-VIS-022 / FORGESIGHT MANUFACTURING
Computer vision became evidence, not theater
A governed vision system inspected 12,000 parts per shift while keeping abstention and human disposition explicit.
ForgeSight Manufacturing
A prototype detected visible anomalies under controlled conditions, but production lighting, part variants, line speed, and human disposition created a different operating problem.
12K parts per shift · 7 governed defect classes · 2023 · 6 months · project
The problem beneath the brief
The model score had become the decision. There was no approved abstention state, no controlled image/evidence record, and no route for inspectors to challenge or correct a result.
- 12K
- parts inspected per shiftproduction acceptance baseline
- 7
- defect classes governedaccepted failure taxonomy
- 3
- abstention reasonshuman-review workflow
Risk constraints
What could not be traded away.
- false acceptance
- line takt time
- variant and lighting drift
- inspector authority
- evidence retention
Findings
What inspection changed.
- confidence thresholds hid class-specific failure
- lighting drift looked like part drift
- retraining data could enter without disposition provenance
Named team and role pattern
The people attached to this engagement.
- Leila Haddad · senior delivery lead
- computer-vision lead
- quality engineer
- controls engineer
- data engineer
- line supervisor
- security reviewer
Architecture
The operating system we installed.
- 01image capture and provenancegolden-set governance
- 02class-aware evaluationhuman override
- 03abstention and human dispositionlighting baseline
- 04drift monitoringmodel/version trace
- 05versioned model releaserollback
Delivery sequence
Four phases. Evidence at every gate.
- 01
Frame
Define the decision, outcome, work products, authority, dependencies, exclusions, and acceptance evidence.
A named sponsor and principal approve the bounded charter. - 02
Assemble
Inspect the operating reality, then assemble named specialists, context, access, controls, and a delivery plan around the actual work.
The client approves the named team, evidence plan, role boundaries, and stop conditions. - 03
Govern
Build and operate the smallest coherent change with versioned decisions, quality evidence, escalation, and acceptance attached.
The integrated state meets the agreed evidence threshold and every material exception has an owner. - 04
Transfer
Rehearse recovery, resolve exceptions, accept the work, remove temporary access, and transfer operating ownership.
The receiving owner signs the handoff with open limits visible.
Complications
Where the plan had to become more honest.
- The best aggregate model was worse on the rare defect class that mattered most.
- A line-speed improvement reduced image quality below the accepted baseline.
Outcomes
What changed—and what the record proves.
- The system abstained rather than forcing a low-evidence classification.
- Inspectors retained disposition authority and every correction fed a governed review queue.
- Release evidence tied model, camera, lighting, line, and part configuration together.
Lessons
What we would carry into the next system.
- Aggregate accuracy can hide the risk class that matters.
- A vision system includes lighting, capture, workflow, and review.
- Human correction is an operating control, not model embarrassment.
Handoff
The engagement ended with an operating owner.
- 01golden-set owner
- 02drift review cadence
- 03inspector escalation
- 04version rollback
- 05correction and retraining governance
Start with the decision
Bring the priority. We will help bound the work.
If the decisions or constraints look familiar, start with the operating reality—not a preselected solution.
Start a conversation.