Manufacturing Quality Control: CV Copilots & Exception Routing

yaml
policy_id: qc-triage-conn-L3-v1.2
plant: NA-Detroit-01
line: L3
sku: CNX-2147A
regions:
  - US-EAST (metadata only)
  - on-prem (images)
owners:
  qe_lead: "sara.lee@company.com"
  production_supervisor: "d.manuel@company.com"
  rework_cell: "cell-7@company.com"
  it_contact: "plant-it@company.com"
thresholds:
  critical:
    defects: ["crack", "misalignment", "missing-pin"]
    confidence: ">=0.88"
    action: "line_stop_if_3_in_5min"
  major:
    defects: ["gate-vest", "short-shot", "flash"]
    confidence: ">=0.80"
    action: "route_qe_review"
  minor:
    defects: ["surface_scuff"]
    confidence: ">=0.70"
    action: "route_rework"
triage:
  routing:
    qe_review_queue: "ServiceNow/QE-Detroit-01"
    rework_kanban: "MES/Kanban/Rework-Cell-7"
    alert_channels:
      - "Teams:#l3-qc-exceptions"
      - "Slack:#plant-detroit-qc"
  slo_minutes:
    qe_ack: 10
    rework_start: 15
    critical_override: 5
  escalation:
    - after: 10
      to: "qe_lead"
    - after: 20
      to: "production_supervisor"
    - after: 30
      to: "plant_manager_oncall"
review:
  human_in_loop:
    critical: "required"
    major: "required if confidence <0.90"
    minor: "optional"
  sampling:
    rate: "5% of auto-clear parts"
    owner: "QE"
  approvals:
    model_change: "qe_lead + it_contact"
    threshold_change: "qe_lead + production_supervisor"
observability:
  metrics:
    - ftq
    - mrb_hold_time
    - exceptions_per_hour
    - false_positive_rate
  error_budgets:
    false_negative_rate: "<=1%"
    alert_on:
      - "3 critical defects in 5 minutes"
      - "slo_breach >10% daily"
security:
  rbac:
    operator: ["view_predictions"]
    qe: ["view_predictions", "override", "label"]
    supervisor: ["view_predictions", "override", "pause_line"]
    it: ["manage_integrations", "export_metadata"]
  data_residency:
    images: "on-prem/NAS/qc-images"
    metadata: "Snowflake/qa_prod_us_east_1"
  retention:
    images_days: 30
    metadata_days: 365
  audit_trail:
    enable: true
    fields: ["part_id", "timestamp", "defect", "confidence", "owner", "action", "override_reason"]

{
  "title": "Manufacturing Quality Control: CV Copilots & Exception Routing",
  "published_date": "2025-12-01",
  "author": {
    "name": "Lisa Patel",
    "role": "Industry Solutions Lead",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "Industry Transformations and Case Studies",
  "key_takeaways": [
    "Start with one high-defect station and define exception owners before model tuning.",
    "Edge-first, governed architecture keeps images in-plant while logging every override and decision.",
    "Expect 3–5 point FTQ lift and 30–40% reduction in MRB hold-time in a sub-30-day pilot.",
    "Use a triage policy YAML to align thresholds, SLOs, and escalation paths across shifts and plants.",
    "Prove ROI with escaped defects down and scrap/hour stabilized; then scale by line, not by plant."
  ],
  "faq": [
    {
      "question": "Will the CV copilot slow the line?",
      "answer": "No. We run edge inference with sub-100ms latency and only pause on configured critical triggers. For all other cases, the copilot routes exceptions asynchronously while the line keeps moving."
    },
    {
      "question": "Do we need data scientists on-site?",
      "answer": "No. Our team tunes models remotely, while your QE defines thresholds in a triage policy. We also provide an enablement kit so operators and line leads can adjust lighting and camera placement safely."
    },
    {
      "question": "How do you handle audits and customer claims?",
      "answer": "Every decision is logged with image links, confidence scores, and owner actions. During an 8D or PPAP review, you pull a time-bounded export with immutable IDs and residency proofs."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "Tier-1 automotive connector supplier with 6 plants (NA/EU), 18 inline cameras, MES: Ignition + SAP ME.",
    "before_state": "FTQ at 93.2% with inconsistent exception ownership; MRB hold-time averaged 84 minutes; three escapes last quarter.",
    "after_state": "FTQ lifted to 97.1% in 28 days; MRB hold-time down to 49 minutes; escapes per million reduced by 62%.",
    "metrics": [
      "First-Time Quality +3.9 points",
      "MRB hold-time -41%",
      "Escapes per million -62%",
      "QE rework hours returned: 240/month across two lines"
    ],
    "governance": "Edge inference kept images on-prem with RBAC aligned to AD groups; all detections and overrides logged with immutable IDs in Snowflake; data residency documented; models never trained on client data; human-in-the-loop for critical calls."
  },
  "summary": "COOs: transform inline inspection with CV copilots and governed exception routing. 30-day audit→pilot→scale plan to lift FTQ and cut hold time."
}