No identification
No facial recognition, no worker identification and no per-worker statistics.
Technology, privacy & security
Privacy by design, and honest about the limits
VISION watches for hazards, not for people. It does not identify workers, keeps evidence only when there is an event, and leaves every decision to a competent person.
Privacy
What VISION does with images
Event-only evidence
Frames are stored only when they carry an event. There is no continuous recording.
Configurable retention
Evidence retention is set per organisation from 7 days to 5 years (90 days by default). Exhibition evidence is deleted automatically, after four hours by default.
Private storage
Evidence sits on private storage and is streamed only after a permission check.
VISION is built around data minimisation: no identification, event-only evidence and configurable retention. We work through Saudi Personal Data Protection Law (PDPL) requirements with your data protection team.
Security
Controls around the data
Role-based access
VISION uses Green Codes Pro’s users, roles and permissions; it is enabled only for organisations that are granted access.
Device controls
Inspector devices are paired, store queued evidence encrypted, and can be revoked and wiped remotely.
Audit trail
Approvals, withdrawals, model-registry changes and device actions are recorded.
Separation of duties
Inspectors cannot approve their own reports; on by default.
AI models
Where the models come from
- Detectors are based on RF-DETR, an open-source detection architecture released under the Apache 2.0 licence.
- Fine-tuning uses openly licensed datasets, with attribution recorded for each one.
- Every model is registered and versioned, and only approved models drive spoken alerts and automatic reporting.
- Each model’s evaluation results are recorded, and we share them during demonstrations.
Limits of AI detection
What AI detection cannot do
- It will miss some real hazards and flag some things that are not hazards.
- It sees an image, not the site: it does not know about permits, controls already in place, or measured distances.
- Performance depends on camera position, lighting, distance and occlusion, and on how closely a site resembles the training data.
- It is not a substitute for supervision, competent persons or any legal duty.
Deployment choices
Decided with you
Where the inference service runs (CPU or GPU, and on whose infrastructure), how long evidence is kept, who can review and approve, and which detectors are switched on are all agreed per project.
See VISION with your own site in mind
Book a guided demonstration. We will walk through the exhibition scenarios, the human review workflow, and what a rollout on your projects would involve.