VisionCrew Inspector

Agentic safety inspection with OpenCV 5 · interactive local demo · synthetic scenes

1 · Pick a scene

or upload

Uploads are resized to 640×480 (the detectors' fixed floor/exit regions assume that frame), processed in memory, and not stored. Detectors are tuned on synthetic scenes; no real-site accuracy is claimed.

2 · What the agent sees

Inspected image
Region chosen for re-analysis
Region chosen for zoom
Crop, upscaled 2× — detector runs again here
Zoomed crop

3 · Agent steps (real trace)

    Raw report JSON
    
  
    

    


    
  
    Perceive → decide → act
    
  
    
    
    Image
    synthetic scene or upload
    →
    
    Perceive · OpenCV 5blur (Laplacian variance)
    exposure (mean luminance)
    spill (HSV colour proxy)
    exit obstruction (Canny edges)
    →
    
    Agent decidesfrom confidence bands
    bounded: 1 reprocess,
    1 zoom per observation
    →
    
    Actverdict + findings
    annotated image
    full decision trace
    
  
    
  
    
    
    Image quality failsREPROCESS once; still failing → needs_human (human review required)
    
    
    Confidence 0.35–0.65 (ambiguous)ZOOM_REANALYZE: crop, upscale 2×, re-detect; still ambiguous → SECOND_PASS
    
    
    Obstruction ≥ 0.80ESCALATE_HUMAN → escalated (human review required)
    
    
    Confirmed ≥ 0.65fail
    
    
    Known limitObstruction 0.65–0.80 → fail with no human-review flag
    
    
    No findingspass
    
  
    
  
    
    
    FastAPI service
    /health, /inspect
    →
    
    AWS Lambda container (ECR)Function URL
    packaged; deployment pending
    →
    
    CloudWatch Logsone structured line
    per inspection
    
  
    
  
    This page runs the same Python pipeline locally. AWS boxes are dashed because the deployment has not been run yet.