A gesture-controlled, fully local AI companion that watches, recognizes, and protects — no cloud, no subscription, nothing leaves the room.
No endless scrolling. Tap any tile and jump directly to that part of the system.
In tradition, Shiva's third eye sees beyond ordinary sight — insight, not just vision. DivyaNetra is built around that same idea: two physical eyes that see, and a third that understands.
Camera A. Watches your hands, not a mouse — hover-and-hold gestures control every part of the HUD.
LIVECamera B. Watches the room — faces, objects, falls — shown live inside the main dashboard.
LIVEA local LLM that turns what the two eyes see into understanding — and decides what to do about it.
IN PROGRESSEverything below is running now on a Jetson Orin Nano — not a mockup, not a concept render.
Hover-and-hold to toggle any feature. No mouse, anywhere in the app.
Detects sustained closed eyes and auto-engages a monitoring mode.
Pose-guided capture — look left, right, up, down. Voice or keyboard, no manual steps.
Local database, cosine-similarity matching — known vs. unknown, instantly.
YOLOv8, 80 object classes, running live on the guardian camera.
Flags a fast upright-to-horizontal posture change and holds the alert.
Local speech-to-text for enrollment — zero cloud dependency.
Live CPU load and temperature, watched on-device in real time.
Migrated end-to-end from an x86 laptop to Jetson Orin Nano.
Each eye runs on its own thread, so a heavy detection model on one never slows down gesture response on the other.
Hand tracking, eye monitoring, HUD control
Shared display, picture-in-picture monitoring panel
Face recognition, objects, fall detection
A real hardware manifest — what's already in hand, and what's needed to finish the vision.
A real build order — completed steps stay lit, upcoming ones are next in line.
Five people, five distinct pieces of the system, one build.
Overall project management, end-to-end fine-tuning, and ownership of the major system components.
Research and analysis — landscape review, competitive study, and the foundation behind the project's direction.
Dataset handling and training — sourcing, preparing, and training the data behind the detection models.
Fine-tuning the end-to-end system — tuning accuracy and performance across the full pipeline.
Hardware configuration and handling — the physical sensing and actuation layer running on real devices.