Flag a banned face. Forget every other one.
GDPR-safe face recognition for UK retail. Recognise a banned individual the moment they enter, at any store across your estate, and whitelist your staff and suppliers so they pass without friction. Everyone else is counted, never identified. And nothing happens to anyone until a human reviews and approves.
Repeat offenders account for a disproportionate share of UK retail theft (BRC trend reporting). QuantumEye delivers real-time AI loss prevention on existing CCTV, ICO-registered and ISO 27001 certified, with human review on every consequential action.
One face graph. Three jobs. Reviewed by a human.
QuantumEye doesn't decide who someone 'is', it surfaces a possible match and asks for human confirmation. That separation is the GDPR-safe path.
Watchlist matching
A face captured at one store is flagged at every other store in your estate. Repeat offenders stop being your other manager's problem.
Whitelist matching
Staff, contractors, VIPs and trusted suppliers pass without triggering alerts. Enrolled in seconds in the dashboard.
Anonymous when not matched
Faces that don't match either list are counted, not identified. No profiles built, no PII stored.
How this is different from live facial recognition.
A shopper who isn't on your watchlist leaves no name, no profile and no record behind. Recognition here is assistive, not automated: it surfaces a known repeat offender to a human, who decides. That is the line between recognition and surveillance.
Vectors live separately from video. Always.
512-dimension embeddings are stored in S3 Vectors (Qdrant). The video itself sits in S3. The two are never co-located inside a row or a clip.
InsightFace at the edge
Face detected on the camera, embedded into a 512-dimension vector, sent to the cloud. The raw face crop is not retained.
Two-index Qdrant search
Banned individuals and whitelist groups are separate vector indices. Match scored, never auto-confirmed.
Human-in-the-loop
A QE Administrator approves the match before any consequential action. No automated ban. No automated police report.
Morning open. A face from other stores in the estate walks in.
Illustrative scenario. The watchlist returns a candidate match; the manager reviews the profile and prior events. Security does a friendly approach. He leaves without incident.
Face Recognition Review
Human review before any action
Nothing happens until a human approves — this match is queued, not actioned.
- Individual
- #QE101
- Status
- Banned
- Source
- Automatic
- Prior matches
- 2 stores
- Added
- 02 May 2026
- Candidate matchWatchlist hit · CAM-02 · entrance14:32:08
- Queued for reviewapproved = false · awaiting human14:32:09
Pairs well with
Shoplifting Detection
Behaviour-based concealment, grab-and-run and till incidents, flagged on the alert track, not in a CCTV review.
PulseWhitelist Groups
Staff, VIPs, suppliers, contractors, granular access, enrolled in seconds in the dashboard.
GuardAfter-Hours Guard
Zero-touch arming. Night-vision detection. Floodlights, siren, evidence pack, all before the call.
Recognised, never identified, without a human saying yes.
The vector is the technical artefact. The profile is the human artefact. QuantumEye never crosses one into the other without explicit approval.
Match ≠ identification
A high-confidence match surfaces a candidate. A human reviewer accepts or rejects before the system attaches a name.
Vectors stored separately from video
Embeddings live in S3 Vectors. Video lives in S3. Never co-located in a DB row or clip metadata.
30-day default retention
Non-banned face data is purged after 30 days. Banned profiles retain until the ban is lifted, plus a grace period.
Right to erasure
Individuals are soft-deleted on request. No hard deletes ever, soft-delete maintains the audit trail.
See it run on your real watchlist.
We can demo against a sample list, anonymised, from your own incident history. The signal is harder to argue with than a marketing example.