The Gesture Eye
Camera A watches your hands. Hover-and-hold gestures control the HUD without a mouse.
LIVEदिव्यनेत्र
DIVYANETRA
Two eyes to see. One mind to understand.
A gesture-controlled, fully local AI companion that watches, recognizes, and protects — no cloud, no subscription, nothing leaves the room.
CLICK / TAP · SWIPE · ARROW KEYS
01 · THE NAME
Why three eyes?
In tradition, Shiva's third eye sees beyond ordinary sight — insight, not just vision. DivyaNetra follows that idea: two physical eyes that see, and a third that understands.
The Guardian Eye
Camera B watches the room — faces, objects and falls — while monitoring stays separate from control.
LIVEThe Inner Eye
A planned local LLM layer will turn what the two eyes see into natural-language understanding and decisions.
IN PROGRESS02 · CAPABILITIES
What it already does
The working prototype runs on Jetson Orin Nano. Tap a capability to open its explanation.
Gesture Control
Hover-and-hold to toggle features without a mouse.
WORKING→Eye-Closure Monitoring
Sustained closed eyes trigger a monitoring mode.
WORKING→Face Enrollment
Pose-guided enrollment with voice or keyboard input.
WORKING→Face Recognition
Local database identifies known vs. unknown people.
WORKING→Object Detection
YOLOv8 detects 80 object classes on the guardian camera.
WORKING→Fall Detection
A v1 heuristic flags a fast upright-to-horizontal transition.
WORKING→Offline Voice Input
Local speech-to-text keeps core interaction cloud-free.
WORKING→System Vigilance
CPU load and temperature are monitored on-device.
WORKING→Edge Architecture
Two camera pipelines run independently to preserve responsiveness.
DEPLOYED→CAPABILITY · 01
Gesture Control
What it does
A mouse-free HUD
The gesture camera continuously tracks interaction so features can be toggled through hover-and-hold gestures. The interface remains responsive while the heavier monitoring pipeline runs separately.
Built for the projector/HUD vision: day-to-day interaction should not require a keyboard, mouse or phone app.
Status
- Working prototype
- Gesture-controlled HUD
- Camera A pipeline
- Local processing
CAPABILITY · 02
Eye-Closure Monitoring
What it does
Detect sustained eye closure
The prototype monitors eye closure and automatically engages a monitoring mode after sustained closure. It is one of the working interaction/safety behaviors already running on the edge device.
Status
- Working
- Automatic monitoring mode
- Local camera processing
CAPABILITY · 03
Face Enrollment
What it does
Pose-guided enrollment
Enrollment guides the person through poses such as looking left, right, up and down. Voice or keyboard input can drive the process, avoiding manual capture steps.
Status
- Working
- Pose-guided capture
- Voice or keyboard driven
- Local database workflow
CAPABILITY · 04
Face Recognition
What it does
Known vs. unknown
InsightFace runs locally with a local database to recognize people and distinguish known from unknown faces. No cloud recognition service is required for the core system.
Technology
- InsightFace
- Local database
- Known / unknown detection
- Jetson Orin Nano
CAPABILITY · 05
Object Detection
What it does
80 classes, on-device
The guardian camera runs YOLOv8 for live object detection across 80 object classes. This is part of the working safety-monitoring pipeline.
Technology
- YOLOv8
- 80 object classes
- Guardian camera
- Local edge inference
CAPABILITY · 06
Fall Detection
What it does
Detect a rapid posture change
The current v1 approach flags a fast upright-to-horizontal transition and holds the alert. It is intentionally described as a heuristic rather than a finished action-recognition model.
Status
- Working v1 heuristic
- Guardian camera
- Safety monitoring
- Future refinement planned
CAPABILITY · 07
Offline Voice Input
What it does
Speech without the cloud
Vosk provides local speech-to-text for the current workflow, keeping core voice input independent of cloud APIs and internet availability.
Technology
- Vosk
- Local speech-to-text
- Zero cloud dependency
CAPABILITY · 08
System Vigilance
What it does
Know the machine is healthy
The system health layer watches live CPU and temperature statistics directly on the device, giving the dashboard awareness of its own operating state.
Status
- Working
- Live CPU statistics
- Temperature monitoring
- On-device
03 · ARCHITECTURE
Two cameras, split by purpose
The system deliberately separates the lightweight gesture-control loop from the heavier monitoring loop, so detection features do not degrade HUD responsiveness.
Gesture
Hand tracking, eye monitoring and HUD control.
Jetson Orin Nano
Local inference, modular capabilities and shared dashboard.
Guardian
Face recognition, object detection and fall detection.
Current local stack includes MediaPipe for hand/body tracking, InsightFace for face recognition and YOLOv8 for object detection. A local LLM is the planned next decision-making layer.
04 · EQUIPMENT
What it runs on
A hardware manifest separating what is already available from the equipment planned for the complete product.
NVIDIA Jetson Orin NanoEdge AI compute — core of the system
IN HAND2× USB WebcamsDedicated gesture and guardian cameras
IN HANDESP32 MicrocontrollerPhysical actuation layer — relays and sensors
NEEDEDFlame / Smoke / Temperature SensorsFire detection fused with camera vision
NEEDEDRelay Modules, Water Pump, Exhaust FanTurning detection into physical response
NEEDEDCompact ProjectorHUD projected into the room itself
NEEDED05 · ROADMAP
What's lit. What's next.
Completed capabilities stay lit; the remaining steps describe the path toward the complete product.
Gesture-Controlled HUD
Hover-and-hold interaction, fully mouse-free.
Face & Object Systems
Enrollment, recognition and 80-class object detection.
Fall Detection & Edge Migration
Heuristic fall detection running on Jetson Orin Nano.
The Inner Eye — Local LLM
Natural-language command and decision-making on-device.
Voice Output & Wake Word
Spoken responses and hands-free activation.
Fire Sensing & Physical Actuation
ESP32 layer — sensors, relays, pump and fan.
Remote Alerts & Projector Display
Telegram notifications and HUD projected into the room.
Fight Detection
Later-stage action-recognition modeling, deliberately scoped as a harder problem.