Live intelligence that watches with you. Sudarshan scales AI detail to your attention: rich, real-time overlays on the feed you are focused on, and a clean, glanceable signal on every other tile. It lets you watch for anything in plain language and drops the alert, in context, the instant it happens.
Named for the Sudarshan Chakra, the ever-vigilant, all-seeing discus, it is built around one rule your operators can feel: the live feed is sacred. Overlays never cover, freeze, or degrade the video. Everything is enforced server-side against your licence, scoped per camera, and runs entirely on your own infrastructure.
12
Live Overlay Models
6
Live Watch Conditions
<1s
Overlay Latency
55+
Languages
Sudarshan is not another box of analytics. It is a live layer over your video wall that puts the most detail where your operators are looking, keeps the rest of the wall calm, and never gets in the way of the feed itself.
AI detail follows your focus. The feed you are watching gets rich, live overlays: bounding boxes, tracks, faces, plates, zones, and lines. Every other tile on the wall stays clean and shows only a quiet signal when something needs you. Intelligence scales to attention, so a 64-camera wall never dissolves into noise.
Overlays render as a precise, translucent layer that never blocks, covers, freezes, or degrades the video. Boxes stay thin, labels stay out of the way, and the picture underneath is always whole. Your operators never lose the ground truth to the graphics drawn on top of it.
Type what matters. "A red truck at the loading dock." "Anyone crossing this line." "This face." "Plate MH12AB1234." Sudarshan arms a live watch and flags the exact moment it happens, right on the wall. Six condition types span faces, plates, plate patterns, objects in a zone, line crossings, and free-text semantic search.
When something fires, the tile rings, the alert rail fills, and one click drills straight into the exact track that triggered it. From there, open the recording at the precise second it happened. No scrubbing through hours of footage to find the moment that mattered.
Every overlay is filtered server-side against your licence before a single byte leaves the backend. Unlicensed model data never reaches the browser, so the boundary can never be bypassed from the client. Per-camera scoping means overlays appear only on the cameras you configured for AI.
Sudarshan scores every alert for dispatch priority and measures which sources actually earn an operator response, then suggests retunes for the noisy ones. Inspired by ISA-18.2 alarm rationalization and AVS-01 dispatch scoring, it turns a firehose of alerts into a ranked, actionable queue.
Describe what matters in a sentence and Sudarshan arms a live watch on the wall. It flags the exact moment your condition is met, not a report you read hours later.
Watch for a red truck at the loading dock
Armed. I will flag any truck-class detection inside the Loading Dock zone and ring that tile the moment one appears. This is a live watch, so it stays active on the wall until you disarm it.
Alert me if anyone crosses the fence line, either direction
Watching Perimeter East. A line-crossing watch is armed on the fence tripwire for both directions. Crossings raise a warning-level signal with the direction of travel and drop a marker on the exact track.
Tell me the instant you read this plate: MH12AB1234
Plate watch armed across 6 ANPR cameras. I will match MH12AB1234 in real time and surface the camera, timestamp, and a snapshot the instant it reads, then hold the alert until acknowledged.
Sudarshan draws boxes for the models that produce them, straight onto the focused feed, from everyday object and face tracking to fire, weapon, and abandoned-object detection. Every overlay is filtered against your licence before it leaves the server, so operators only ever see what they are entitled to.
Live bounding boxes and stable track IDs for people, vehicles, and objects as they move through the frame.
Stable per-person track IDs that follow the same individual across frames, so an operator can keep a target locked as it moves.
Identity overlays on watch-listed faces, with confidence, drawn only where face recognition is licensed and configured.
Number plates read and rendered in real time, powering plate watches and plate-pattern matches.
Skeletal keypoints for posture and activity, the foundation for fall, loitering, and behaviour signals.
Per-person safety-gear compliance drawn on the feed, flagging anyone missing required protective equipment in the zone.
Named zones that flag entry, dwell, and occupancy the moment a target enters the area you drew.
Directional tripwires that light up the instant a track crosses, with the direction of travel.
People-count and density overlays that turn a busy scene into a number you can watch and threshold.
Fire and smoke boxed live on the focused feed the instant the detector sees them, when every second counts.
Gun and knife boxes, shown only after temporal confirmation so a single-frame false positive never flashes on the wall.
Unattended bags and packages boxed once they dwell past your threshold with no owner nearby.
A gunshot, a scene that stops looking normal, a blinded camera, a tailgate at a door: none of these are a bounding box you can draw. Sudarshan surfaces every one of them on the same wall through the alert lane, with a tile ring, a rail entry, and a one-click drill-in, so nothing important is silent just because it cannot be outlined.
Gunshot, glass-break, and scream detection. Sound has no bounding box, so Sudarshan rings the tile and fills the alert rail the instant it is heard.
A self-learning, scene-level score for "this does not look normal". Not a box, but a signal on the wall the moment a scene deviates from its baseline.
Defocus, blackout, and obstruction flip the tile to alert, so a blinded camera is never mistaken for a quiet one.
Piggybacking through a controlled door raises a signal on the wall tied to the exact entry event.
An alert on the wall is only useful if it takes you somewhere. Sudarshan makes every signal a doorway into the moment it describes.
The tile that fired gets a coloured ring and rises in the alert rail. Watch hits carry their own distinct signal, so a thing you personally asked to see never blends into the routine noise.
One click expands the camera and jumps to the exact track that triggered the alert, highlighted on the live feed with its identity, plate, or class already resolved.
From the alert, open the recording at the precise timestamp it happened. No scrubbing, no guessing. The evidence is one motion away from the signal.
Live intelligence must never become a way to leak what an operator is not entitled to see. Sudarshan enforces every boundary on the server, where it cannot be bypassed.
The backend strips every unlicensed model from each overlay frame before it is sent. An operator without a face-recognition licence never receives face data, even in the raw stream. The gate lives on the server, where it cannot be tampered with from the browser.
Overlays appear only on cameras you explicitly configured for AI. A camera with no AI models produces no overlays and no watch signals, so privacy-sensitive areas stay untouched by inference.
Sudarshan is designed around a single rule: the live video is sacred. Overlays are a thin translucent layer, never a modal, never a freeze, never a cover. If overlay data goes stale, it fades out and the untouched feed remains.
All inference, watch evaluation, and overlay assembly run inside your own infrastructure. No frames, no detections, and no watches are sent to any external service. Sudarshan works in fully air-gapped deployments without modification.
Most video walls either paint boxes on every tile until the screen is unreadable, or bury AI behind a separate analytics tab. Sudarshan puts live intelligence exactly where the operator is looking, and a quiet, glanceable signal everywhere else.
Watches are live and spoken in plain language. This is not a saved search you run after the fact. It is an armed watch that flags the instant it happens, on the same wall your operators already stare at.
The live feed is treated as sacred. Overlays are a translucent layer that never covers, freezes, or degrades the video, so the picture underneath is always trustworthy.
Licence enforcement happens on the server. Unlicensed model data is never sent to the browser, so the boundary cannot be defeated from the client or by inspecting network traffic.
Every alert carries a dispatch-priority score and a one-click path to the exact recorded second it fired, turning "something happened" into "here is the footage" in a single motion.
Schedule a demo and watch attention-scaled intelligence, live plain-language watches, and in-context alerts running on a real multi-camera wall.