GuardVision Edge
Turn spare Android phones into privacy-first AI guardians, then scale the same behavior-based engine into commercial spaces. GuardVision Edge analyzes anonymous tracks, zones, motion, pose, and object proximity on-device, then surfaces explainable alerts for human review.
- Launch path
- Spare phone
- Runtime
- Edge-only
- Review model
- Human-led
Live Risk Console
Android device - Entry and counter view
Explainable alert
Track #014 remained near a sensitive zone for 180 seconds and crossed the limit line 4 times. The system reports behavior evidence, not identity.
Track #014
HighSensitive zone dwell
Risk score 72
Track #021
HighRapid approach to counter
Risk score 64
Track #008
MediumObject near right hand
Risk score 48
Perception and behavior AI
Signals built for reviewable risk, not identity surveillance.
The product story now combines a consumer Android app entry point with an enterprise-ready safety platform, while keeping the same behavior-based privacy boundary.
15 FPS target
Anonymous track intelligence
Person detection creates short-lived Track IDs for motion, dwell time, zone entry, and crowd context without identity profiles.
33 landmarks
Pose-aware behavior signals
Pose landmarks and temporal sequences help separate routine movement from fast approach, falls, pushing, and raised-arm events.
Multi-cue
Object proximity reasoning
Object detections become weak signals until correlated with hands, bodies, dwell windows, or unattended-area rules.
AI capability map
A layered perception stack, presented as operator evidence.
GuardVision Edge combines detection, pose, object, zone, and scoring signals. Each capability is framed as reviewable evidence instead of an identity or intent claim.
EfficientDet-Lite person boxes
Finds people in the frame, normalizes bounding boxes, and starts anonymous Track IDs for short-lived behavior analysis.
Signals
- Bounding box
- Confidence
- Frame timestamp
- Track seed
Output
Anonymous person observations for tracking.
MediaPipe pose landmarks
Extracts body landmarks and temporal motion cues to support raised-arm, push, fall, and fast movement review.
Signals
- 33 landmarks
- 3D world coordinates
- Motion direction
- Pose quality
Output
Pose features attached to an anonymous Track.
Suspicious object proximity
Treats object detections as weak signals until they are correlated with hands, bodies, dwell windows, or unattended-area rules.
Signals
- Object class
- Object box
- Hand proximity
- Duration
Output
Suspicious object signal for human review.
Geofenced behavior context
Maps tracks to configured zones and limit lines so dwell time, crossings, and restricted-area proximity can be explained.
Signals
- Zone polygon
- Line crossing
- Dwell time
- Entry and exit
Output
Spatial behavior evidence, not identity.
Multi-cue confidence model
Combines time, space, motion, pose, and object cues so no single weak signal becomes a high-risk alert by itself.
Signals
- Event weight
- Model confidence
- Time factor
- Zone sensitivity
Output
Explainable score band for operator triage.
Review-ready event reasons
Formats risk output into track, zone, duration, evidence, and recommended review text for audit-friendly operations.
Signals
- Reason list
- Risk level
- Evidence summary
- Review status
Output
Human-readable alert context.
Frame to risk feature pipeline
Every model output is normalized before reaching the analysis layer, keeping future modules compatible with the same product surface.
01
Frame is resized, rotated, and mapped into model space.
02
Model outputs become normalized detections, poses, and object signals.
03
Track and zone context convert raw outputs into behavior features.
04
Risk engine generates score, reason list, and review recommendation.
Capability boundaries
The feature section keeps product claims aligned with the system design.
- No facial identity matching.
- No protected-attribute scoring.
- No automatic punishment or law-enforcement action.
- High-risk alerts require multi-cue evidence and operator review.
Edge AI and privacy-first security
The invisible AI guardian inside an Android phone.
GuardVision Edge explains edge computing through a concrete safety product: camera frames are interpreted by the phone itself, sensitive video does not need to travel to a cloud recognition service, and alerts describe observable behavior instead of personal identity.
0
Cloud video recognition required by default
On-device
Runtime location for inference
None
Identity database created
A phone becomes a local AI guard post.
The device is no longer only a screen. CameraX, LiteRT, and the Android runtime turn it into a real-time security computer that can reason about time, space, movement, and object proximity before sending any sensitive media elsewhere.
Behavior
Abnormal movement
Repeated paths, irregular motion, or sudden direction changes become reviewable movement evidence.
Approach
Dangerous approach
Fast movement toward a counter, door, cash area, or protected zone can raise the score when other cues agree.
Dwell
Long stay near sensitive areas
The local engine measures how long an anonymous Track remains inside a configured zone.
Interaction
Risky object relationship
Object detections stay weak until correlated with a hand, body, zone, or unattended-duration rule.
Privacy-first by architecture, not by marketing copy.
The product boundary is simple: understand what is happening in a space without collecting faces, creating identity profiles, or depending on cloud recognition.
- On-device inference keeps image analysis close to the camera source.
- No facial recognition, face embeddings, or protected-attribute scoring.
- Anonymous Track IDs are short-lived and scoped to behavior review.
- Cloud services are not required for local risk detection or alert reasoning.
Cloud AI versus edge AI
The difference is not only speed. It changes who controls the sensitive data path.
Android Edge AI Pipeline
The runtime section gives technical buyers a concise map from camera frame to explainable alert.
- 01CameraX ImageAnalysis keeps the newest frame and avoids delayed stale-frame decisions.
- 02Frame preprocessing maps rotation, scale, and model input coordinates consistently.
- 03LiteRT / TensorFlow Lite runs person, pose, or object models locally with CPU, GPU, or NPU acceleration paths.
- 04Anonymous Track IDs, zone rules, dwell timers, and motion vectors become behavior features.
- 05The risk engine produces a score band, reason list, and local alert for human review.
Invisible Guardian: Edge AI, Android runtime, and privacy safety.
This section is written as a durable entry point for technical education, SaaS dashboard storytelling, sales enablement, and a future official whitepaper about privacy-first security AI.
- Edge computing fundamentals
- Mobile on-device AI runtime
- Privacy-safe behavior detection
Privacy-first behavior detection
Inside GuardVision Edge: how AI sees risk without recognizing people.
GuardVision Edge does not ask who someone is. It de-identifies camera frames into anonymous coordinates, temporary Track IDs, motion vectors, pose landmarks, zone events, and explainable risk scores.
From 'who are you' to 'what is happening'
Traditional surveillance starts with face matching and identity databases. GuardVision Edge treats personal appearance as out of scope and digitizes the judgment a trained guard would make from observable behavior: loitering, rapid approach, intrusion, crowding, or dangerous object interaction.
Visual pipeline: pixels to privacy-safe data
CameraX captures the newest frame, LiteRT runs local inference, and the app keeps only the structured observations needed for behavior analysis. Raw frames are not turned into identity profiles.
Input
RGB camera pixels
AI output
Anonymous bounding boxes
Person shape is located without storing a face embedding.
Input
Eyes, nose, and facial traits
AI output
33 pose landmarks
Pose becomes a digital stick figure for motion reasoning.
Input
Background and objects
AI output
Zone coordinates and object proximity
Sensitive areas and object interactions become spatial signals.
De-identification: turning people into privacy-safe data twins
The website now explains how high-privacy image content is reduced into geometry, short-lived IDs, and skeleton keypoints before the risk engine sees it.
Module
Person detection
Original image content
Facial expression and clothing detail
Anonymous AI data
Bounding box center and size
Privacy guardrail
Geometry does not contain face texture or biometric embeddings.
Purpose Define physical coverage in the frame.
Module
Anonymous tracking
Original image content
Name, account, or identity record
Anonymous AI data
Random session Track ID such as Track #001
Privacy guardrail
The ID is temporary and cannot be used as a cross-day identity profile.
Purpose Follow the same target during one live session.
Module
Pose estimation
Original image content
Body appearance, skin tone, or personal traits
Anonymous AI data
33 skeleton keypoints and 3D pose coordinates
Privacy guardrail
Only joint relationships are kept for behavior reasoning.
Purpose Recognize motion patterns such as falling, pushing, or striking.
Temporary Track IDs replace identity profiles
Each person becomes a session-scoped Track ID such as Track #101. The ID supports motion analysis but is not connected to a name, face, account, or identity database.
Center point
Current XY position in frame space
Motion vector
Direction and approach trend
Instant speed
Normal walking versus abnormal sprinting
Dwell timer
Time spent inside a configured sensitive zone
Three risk sensors power the behavior model
The analysis engine combines movement, time, and space. Each signal is transparent enough to be inspected by an operator.
01 / Track geometry
Direction, velocity vector, path recurrence, and route regularity identify abnormal movement and dangerous approach.
Path anomaly = velocity threshold + recurrence threshold
02 / Dwell time
The engine measures how long a Track stays still or repeatedly returns to a configured area such as an ATM, storage door, or counter.
Dwell risk = dwell seconds / allowed time x base weight
03 / Geofencing
The center of a bounding box is checked against virtual lines and polygonal sensitive zones.
Zone risk = in sensitive zone x zone priority factor
Behavior logic uses transparent conditions
The risk engine checks observable events instead of guessing intent. Every alert can be explained as a combination of time, space, motion, pose, and object evidence.
- OKDid a track overlap a polygonal sensitive zone beyond the threshold?
- OKDid arm landmarks show a fast strike-like trajectory?
- OKDid two Track IDs move with suspiciously matched vectors?
- OKDid the density of Track IDs increase sharply in a small area?
- OKDid an object remain close to a hand or unattended for long enough?
Risk score: multi-cue evidence, not black-box judgment
A single behavior is not treated as guilt. The system combines weak signals into a stronger alert only when space, time, speed, pose, and object cues support the same review decision.
0-30
Low
Record metadata without interrupting the scene.
31-60
Medium
Mark the frame and write an event log.
61-80
High
Send a strong-signal alert for review.
81-100
Critical
Save the evidence window and escalate the alert.
Responsible AI takeaways
01
Edge AI keeps inference on the Android device and avoids cloud video analysis by default.
02
Weighted logic makes alerts explainable as accumulated evidence, not hidden identity scoring.
03
Anonymous pose, motion, and zone analysis can support safety without turning the product into face surveillance.
Live risk workflow
From pixels to operator-ready evidence.
- 01CameraX ImageAnalysis keeps the newest frame and avoids delayed alerts on Android devices.
- 02LiteRT and ML Kit models convert pixels into detections, pose landmarks, and object signals.
- 03Analysis rules combine dwell time, zones, motion, and multi-cue confidence without facial identity.
- 04Alerts include reasons, timestamps, and anonymous Track IDs for human review.
Market position
Start with spare phones. Expand into professional spaces.
GuardVision Edge follows a dual-track go-to-market path: a consumer app that reuses idle Android devices for personal space protection, and a long-term B2B platform for offices, warehouses, schools, and public spaces.
The commercial value is not knowing who someone is. It is helping people understand what is happening in a space.
Personal space guardian
A fast Google Play entry point for renters, solo living, small shops, and temporary travel setups.
- Spare Android phone reuse
- Simple zone guard setup
- On-device alerts and local evidence
- Freemium path into Plus and Pro
Commercial safety platform
The same edge AI core expands into managed deployments for commercial and public environments.
- Camera placement and zone calibration
- SLA and device licensing
- White-label or OEM channel options
- Operator review and audit workflows
Technical architecture
Android edge AI stack for low-latency privacy protection.
The website now reflects the presentation architecture: CameraX for frame acquisition, LiteRT for local inference, ML Kit or MediaPipe for pose and object signals, and encrypted local storage for event metadata.
Capture
Android CameraX ImageAnalysis
Backpressure control keeps the stream current and avoids stale alerts.
AI runtime
LiteRT / TensorFlow Lite
GPU and NPU acceleration keep inference local, efficient, and offline-ready.
Perception
ML Kit / MediaPipe
Pose, object, and tracking signals support reviewable behavior evidence.
Storage
Room / SQLite with SQLCipher path
Event metadata and evidence can stay encrypted with retention controls.
Anonymous tracking
Bounding boxes become temporary Track IDs for motion, dwell, and zone reasoning within a session.
Behavior feature extraction
Speed, direction, dwell time, zone relationships, pose quality, and object proximity become structured features.
Risk scoring engine
Multi-cue rules convert behavior features into a 0-100 score with a reason list and review recommendation.
Overlay and dashboard UI
Bounding boxes, virtual zones, alert labels, and risk cards create an inspectable live dashboard.
Behavior risk catalog
Observable events the system can explain.
Risk is derived from time, space, motion, pose, and object relationships. A single weak signal is never presented as a definitive high-risk conclusion.
Commercial model
Freemium app entry, enterprise deployment path.
The consumer app validates demand quickly through Google Play. The enterprise path monetizes site setup, per-device licensing, maintenance, and partner channels without changing the privacy-first AI core.
Free
NT$0
Single-device guard mode for lightweight personal monitoring.
- Basic motion detection
- One Android device
- 24-hour local history
Plus
NT$79/mo
AI person detection and custom zone protection for renters and small spaces.
- Anonymous person detection
- Custom guard zones
- 7-day encrypted event history
Pro
NT$179/mo
Advanced AI events and multi-device readiness for families and small businesses.
- Pose and object event tiers
- Multi-device support
- 30-day event retention policy
B2B expansion
Commercial deployments add field calibration, rule tuning, licensing, SLA maintenance, and partner-led installation.
Product principles
The website language now mirrors the safety doctrine.
Risk should be based on verifiable behavior patterns and environmental events, not appearance or facial traits.
The system output should not be 'this is a dangerous person.' It should say 'an anonymous track stayed in a sensitive zone for more than 180 seconds.'
Security apps are built on trust. Every alert needs explainable evidence.
Reusing an old phone lowers hardware cost and gives idle devices a new safety value.
Privacy-preserving by default
Behavior first. Identity out of scope.
The public privacy page documents what the product does, what remains out of scope, and which evidence operators must review before action.
Deployment paths
Consumer-first, enterprise-expandable.
The same Android edge AI core supports spare-phone personal protection first, then expands into commercial pilots and managed deployments.
FAQ
Product boundaries operators can trust.
Does GuardVision Edge identify people?
No. The product is designed around anonymous tracks, behavior events, and local processing instead of facial identity or biometric matching.
Can it automatically decide someone is dangerous?
No. Alerts explain observable behavior and ask operators to review context before taking action.
Is Google Play available now?
The public Android listing is planned. Pilot and enterprise deployment discussions can start through the contact flow.
Launch and contact
Choose the next step without implying a public release.
The public Google Play listing is not live yet. Operators and enterprise buyers can still join the waitlist, contact sales, or request a deployment review.
Public listing coming soon
The Android app entry is reserved for the official release path. Until the listing is live, the website keeps the Play action informative and non-clickable.
The site will not link to a placeholder Play Store URL before a real listing is available.
Early access
Join the pilot waitlist
Get notified when the Android listing, pilot program, or private evaluation track opens.
Join waitlistCommercial
Contact sales
Start a deployment conversation covering site goals, privacy expectations, device profile, and review workflow.
Contact salesField readiness
Request deployment review
Prepare a pilot assessment for camera placement, risk zones, runtime constraints, and operator review policy.
Request reviewCTA guardrails
- No fake Play Store URL before publication.
- No promise of automatic enforcement or identity recognition.
- Pilot and deployment requests require human review.