Technical concept

Architecture & specification

A modular monolith that runs on a Windows CPU, offline, next to the line. It writes nothing to the PLC in V1 and never auto-passes an image it cannot trust.

System context

From photon to audit trail

FACTORY NETWORK — NO PLC WRITES IN V1

USB / industrial camera

Fixed mount, controlled light

Capture & quality check

Exposure, focus, brightness band

Registration / alignment

Fixture fiducials, pose correction

Recipe-based feature checks

OpenCV baseline · later anomaly-detection research track

Decision engine

PASS / FAIL / REVIEW

Evidence store

Local images + SHA-256 hashes · SQLite → PostgreSQL

Operator dashboard (React)

Human review, append-only

Reports / CSV

Evaluation & audit exports

Flow reads left→right, top→bottom

Stack

Recommended stack

Frontend	React + TypeScript + Vite
Backend	Python FastAPI, versioned under /api/v1
Database	SQLite (dev) → PostgreSQL
Vision	OpenCV + NumPy; later PyTorch / anomaly methods (e.g. PatchCore-style) as research
Runtime	Windows, CPU-only, offline

Entity model

Product

id, name, customer_ref, active

ProductVersion / Recipe

id, product_id, version, checks[] (JSON), created_by, created_at

Image

id, source_type SYNTHETIC | LAB_PHOTO | CUSTOMER_AUTHORISED, sha256, provenance, captured_at

InspectionResult

immutable · image_id, recipe/model version, decision, evidence, timestamp

HumanReview

append-only · result_id, reviewer, decision, comment, timestamp

AuditEvent

actor, action, entity, payload hash, timestamp

Recipe

Versioned, human-readable checks

recipes/housing-A@1.2.0.json (excerpt)json

{
  "id": "housing-A",
  "version": "1.2.0",
  "product": "Synthetic plastic housing (demo)",
  "review_margin": 0.15,
  "checks": [
    {
      "id": "top_clip",
      "description": "Teal retaining clip seated in upper socket",
      "roi": {
        "x0": 184,
        "y0": 139,
        "x1": 306,
        "y1": 177
      },
      "hsv": {
        "hRanges": [
          [
            160,
            190
          ]
        ],
        "sMin": 0.27,
        "vMin": 0.24
      },
      "mode": "present",
      "threshold": 0.4
    },
    {
      "id": "lower_clip",
      "description": "Teal retaining clip seated in lower socket",
      "roi": {
        "x0": 184,
        "y0": 249,
        "x1": 306,
        "y1": 288
      },
      "hsv": {
        "hRanges": [
          [
            160,
            190
          ]
        ],
        "sMin": 0.27,
        "vMin": 0.24
      },
      "mode": "present",
      "threshold": 0.37
    }
  ]
}

API

Local REST endpoints

POST/api/v1/inspections
GET/api/v1/inspections/{id}
POST/api/v1/inspections/{id}/reviews
GET/POST/api/v1/products
GET/POST/api/v1/products/{id}/recipes
GET/api/v1/reports/evaluation

decision.pypython

def decide_inspection(capture_quality, registration, feature_checks, policy):
    if not capture_quality.ok or not registration.ok:
        return Decision.REVIEW, [class="s">'Image unsuitable for reliable inspection']
    if any(check.status == class="s">'FAIL' for check in feature_checks):
        return Decision.FAIL, failed_check_reasons(feature_checks)
    if any(check.status in (class="s">'UNKNOWN', class="s">'UNCERTAIN') for check in feature_checks):
        return Decision.REVIEW, uncertain_check_reasons(feature_checks)
    if not feature_checks:
        return Decision.REVIEW, [class="s">'No enabled checks in recipe']
    return Decision.PASS, [class="s">'All configured visible feature checks passed']

Engineering rules

Eight hard rules

1.Every event carries source type, recipe version, timestamp and evidence.
2.Out-of-focus, under/overexposed or unregistered images never auto-PASS.
3.Human review appends; it never overwrites a result.
4.Offline is mandatory; no cloud upload without explicit approval.
5.No accuracy, revenue or funding claims without evidence.
6.This is not a machine-safety certification.
7.Modular monolith — no premature microservices.
8.CPU-only; no GPU dependency.

Research questions & work packages

Compare

OpenCV recipe baseline
Anomaly detection (research track)
Recipe + human review workflow

Measure

Recall, false rejects, escaped defects
Review rate and cycle time
Variation across lighting and position
Changeover hours per new variant

Plan

Phased development

Security & privacy

Customer images stay put

GDPR
Avoid capturing people; document any personal data processing.
IP of customer images
Images belong to the customer; usage defined in a data-rights agreement.
Local storage
Evidence stored on-premises with SHA-256 integrity hashes.
Retention / deletion
Configurable retention with verifiable deletion.
Roles
Operator, quality manager, administrator — least privilege.