ProtocolSoftware // Optical & Cognitive QA Layer

PrimeVision: Automated Gemini 3.1 Pro visual & linguistic auditing.

Decoupled from test execution, PrimeVision is a stateless optical verification engine combining OpenCV bounding-box geometry with Gemini 3.1 Pro Vision. Ingests raw screenshot matrices from PrimeCapture or CI/CD pipelines to detect UI clipping, text overflows, RTL mirroring bugs, and contextual localization errors.

Inspect Audit Pipeline ↓ View Upstream /capture →

vision-engine // audit-report.json

// Multimodal Audit Payload (de-DE / step_04.png)
{
  "engine": "prime-vision-v1",
  "model": "gemini-3.1-pro-vision",
  "locale": "de-DE",
  "opencv_bbox_collision": true,
  "defects_found": [
    {
      "severity": "CRITICAL",
      "type": "TEXT_OVERFLOW_CLIPPING",
      "string": "Kontoeinstellungen speichern...",
      "bounds": [142, 880, 938, 964]
    }
  ],
  "unlocalized_tokens": 0
}

Three-Stage Optical & Cognitive Pipeline

ARCHITECTURE // STATELESS API

Stage 01 // Spatial Geometry

OpenCV Contour & Bounds

Pre-processes raw UI screenshots using high-speed OpenCV edge detection to map container boundaries, padding tolerances, and overlapping UI elements before invoking LLM tokens.

Stage 02 // Multimodal AI

Gemini 3.1 Pro Linguistic QA

Evaluates in-context translation fidelity, detects hardcoded en-US bleed-through strings, flags broken bidirectional (RTL) layouts in Arabic/Hebrew, and scores cultural appropriateness.

Stage 03 // Triage Output

Structured JSON & Heatmaps

Outputs annotated visual heatmaps highlighting exact defect coordinates alongside strict JSON schemas ready for ingestion into PrimeTrack Kanban bug queues.