Sudarshan puts live AI on the feed your operator is watching and turns detections into rules that decide and escalate on their own. Radha runs the whole system in plain language. Both on your hardware, 100% on-premise, air-gap compatible, and never priced per camera.
Everything you need to ingest, record, analyse and act on video at enterprise scale. Sudarshan watches every feed and runs the automation, Radha answers and operates in plain language, and all 22 AI analytics run on your own hardware.
Detection, recognition and behaviour analytics covering people, vehicles, objects and audio, from ANPR and PPE to intrusion, tailgating, queue analytics and fire detection. All 22 are developed in-house, so there are no plugins, no third-party AI licences and no per-stream AI fees; you scope which analytics you enable to your plan. The per-camera anomaly and tamper models learn each view and refresh automatically.
Sub-second WebRTC latency with native support for RTSP, RTMP/RTMPS, HLS/LL-HLS, SRT, NDI, RIST, ONVIF, and GB28181. Every protocol is implemented in a first-party engine, not an FFmpeg or SRS wrapper.
Direct bitstream to faststart MP4 without transcoding. Native H.265 recording preserving 50% bandwidth savings. Four modes: continuous, scheduled, event-triggered, and manual with timeline playback up to 8x speed.
Access recordings from 15+ brands including OEMs, Hikvision, Dahua, Uniview, Axis, Bosch, Hanwha, and any ONVIF Profile G device. Time-range search, multi-channel forensic grid, and SHA-256 verified downloads.
A conversational, action-taking, fully on-premise VMS copilot. 100+ tools: search video by description, run multi-step investigations with a replayable evidence timeline, diagnose why a camera is down, act on hundreds of cameras at once, size the next server, play back from the NVR, enrol faces, author automations, get shift briefings. It sees and listens, entirely on your hardware.
Live AI on the wall itself. 12 models draw real-time overlays on the focused feed while the rest of the grid stays clean, and you arm a watch by describing it: "a red truck at the loading dock", "this plate", "anyone crossing this line".
Sudarshan Rules combine AND/OR conditions and ordered temporal sequences with branching actions per condition, correlate video against access control, and dry-run against real history before you arm them. Alerts are scored for dispatch priority, not just raised.
On-premise, cloud, hybrid, or edge. Air-gap compatible with zero cloud dependency. Thousands of streams per node, 200+ REST API endpoints, white-label branding, and 55+ language support including RTL scripts (Arabic, Hebrew, Persian, Urdu) and 13 Indian languages.
Sudarshan is live AI on the wall itself. It puts real-time intelligence on the feed your operator is actually watching, arms watches you describe in plain language, and turns detections into automation rules that decide and escalate on their own. Included with any AI plan, with no separate SKU.
The feed you are watching gets rich live overlays: boxes, tracks, faces, plates, and tripwires across 12 models. Every other tile stays clean and shows a quiet signal only when it needs you. A 64-camera wall never dissolves into noise.
Type what matters. "A red truck at the loading dock." "Anyone crossing this line." "Plate MH12AB1234." Sudarshan arms a live watch across six condition types spanning faces, plates, plate patterns, objects in a zone, line crossings, and free-text semantic search.
Automation rules combine AND/OR conditions and ordered temporal sequences, with branching actions per condition, so "badge swipe with no matching face within 10 seconds" becomes a rule rather than a hope. Video correlates directly with access control.
Preview a rule against real history to see exactly what it would have fired on before it ever goes live, then watch it on a rule health board. Physical actions notify your team; Sudarshan alerts and escalates, it does not actuate your hardware.
Every alert gets a dispatch score, and an effectiveness board shows which sources actually earn an operator response, with one-click bounded retune for the noisy ones. Inspired by AVS-01 dispatch scoring and ISA-18.2 rationalization, not a certified implementation of either.
Overlays are filtered server-side against your licence before a byte reaches the browser, so unlicensed model data never leaves the backend and the boundary cannot be bypassed from the client. Per-camera scoping keeps AI on exactly the cameras you chose.
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 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, raising a warning-level signal with the direction of travel.
Tell me the instant you read plate MH12AB1234
Plate watch armed across 6 ANPR cameras. I will surface the camera, timestamp, and a snapshot the instant it reads, then hold the alert until acknowledged.
Ask in plain language and Radha operates your cameras, recordings, alerts, and analytics for you. It runs entirely on your own hardware, answers only from your live system, and replies in under a second. It is the kind of copilot no other video platform puts in your hands.
Ask in plain language and Radha executes. It starts recordings, pulls footage, plays streams, diagnoses cameras that have gone down, acts across the whole fleet at once, and works the alarm console, across 100+ real tools. It does the work, not just the talking.
Say what you mean. Radha understands follow ups, remembers context, learns your site’s camera nicknames, and takes spoken questions through fully local voice input. No commands to memorize, no menus to hunt through.
Radha never makes things up. Every answer comes straight from your live cameras, recordings, and system health. For open questions it runs a real multi-step investigation and returns findings with a replayable evidence timeline. It can even look at a camera and describe the scene.
The model stays warm and resident on your hardware, so answers come back in under a second. Radha is ready the moment you are, every single time.
A purpose-built on-premise language model running entirely on your own infrastructure. Your video and conversations never leave the building. Fully air gap ready with zero cloud dependency.
Radha cannot delete anything. State-changing actions require explicit confirmation naming the exact target. Prompt injection defense, role based access control on every tool, per user rate limiting, and complete audit logging of every action Radha takes.
Why is the loading dock camera offline?
The host answers on the network but the RTSP handshake is being refused, and the last 12 stream errors are all authentication failures. That pattern is a changed camera password rather than a network fault. I can restart the stream once the credentials are updated.
Start recording on Camera 12
Recording started on Camera 12. I will confirm the file size and duration when you stop the session.
Show me crowd alerts from today
Found 4 crowd detection alerts today: Cafeteria at 12:34 PM (78 people), Main Lobby at 9:02 AM (63 people), Auditorium at 2:15 PM (124 people). I can pull up the footage for any of these.
How many more cameras can this server take?
On current headroom, roughly 40 more for recording only, or about 12 more if each one runs a single AI model. Two or more models per camera would need a second GPU. I can break that down per resource if useful.
AI is only half of it. Visylix ships the day-to-day machinery a security team actually runs a site on, so nothing has to be bolted on from a second vendor.
Autotracking runs as a tested control policy, not a demo, alongside preset recall, VMS-driven tours, and a tour editor your operators can edit themselves.
Floor-plan maps with live camera status and alarm badges, plus a geographic map view on your own supplied tiles. Air-gapped sites keep their maps.
A real operator console for acknowledging, escalating, and closing out incidents, with each alert carrying the dispatch score Sudarshan assigned it.
Export an incident as a self-contained, offline-verifiable player. It opens with no Visylix install and proves its own integrity, built for police, prisons, and court submission.
Exact plate journeys with travel times between cameras, plus appearance-based person journeys for narrowing a search. Appearance matching is a lead, not an identification.
Active Directory and LDAP authentication configurable from Settings, with role-based access control across more than 30 permissions and full audit trails.
ONVIF vendor device packs for the top five brands, bulk device administration, camera I/O, and PullPoint event ingestion. Client-side Profile S, T, and G operations are implemented.
An in-product device-health console surfaces failing cameras, storage runway, and service state before an operator discovers the gap during an investigation.
Specialized AI video analytics solutions designed for the unique requirements of smart cities, enterprise security, retail, healthcare, transportation, manufacturing, education, and banking sectors.
Built from the ground up for reliability, security, and scale, with on-premise data custody, role-based access control, and a contractual uptime commitment on Enterprise plans.
TLS 1.2+ enforcement, DTLS-SRTP encrypted WebRTC, AES-256 credential storage, RBAC with 30+ permissions, comprehensive audit trails, and content security policy, built for mission-critical video infrastructure.
Visylix never phones home. After installation the whole system, including the AI and the copilot, operates with no internet connection at all, so an air-gapped site is a supported deployment rather than a degraded one. Enterprise plans add a contractual response commitment.
Redundant containerized microservices with operator-initiated standby cutover, graceful shutdown, leader election, and continuous health monitoring. Built for thousands of concurrent streams per node, sized and validated on your own hardware during the proof of concept.
Face recognition, ANPR/ALPR, object detection, person tracking, crowd detection, PPE detection, heat maps, motion detection, pose estimation, unique person counting, intrusion detection, line crossing, camera tampering detection, abandoned object detection, tailgating detection, queue and wait-time analytics, parking occupancy, speed estimation, anomaly detection, fire and smoke detection, weapon detection, and audio event detection. Every one is developed in-house, so there are no plugins, no third-party AI licences and no per-stream AI fees. You scope which analytics you enable to your plan. Model adaptation is per camera and bounded, on-premise, and operator-supervised.
Visylix is designed for the operating conditions of manufacturing floors, smart city command centres, hospitals, and transport hubs: continuous streams, on-premise data custody, and analytics that run without an operator watching every feed.
22
In-House AI Analytics
8
Industries Served
13
Streaming Protocols
55+
Languages Supported
Start your trial in minutes, no sales call required. Deploy on your own infrastructure and scale from a single location to a multi-site rollout, with no per-camera fees on any paid plan.
The questions we are asked most. Hundreds more are answered on the full FAQ page.
Visylix is automated AI based video surveillance: the system watches, decides, and acts, rather than just recording for someone to review later. It is an enterprise-grade AI video management system (VMS) built by Aptibit Technologies, combining a proprietary native streaming engine, 22 AI analytics, and Radha, the only VMS copilot we know of that is conversational, action-taking and fully on-premise at once. It supports large-scale concurrent connections with sub-second WebRTC live view, up to sub-200ms in ideal conditions, 13 streaming protocols, 200+ REST API endpoints, and serves as a modern alternative to traditional CCTV, DVR, and NVR systems.
There is no per-camera limit and no per-camera licence. The engine is built for thousands of concurrent streams per node on commodity hardware, and the exact per-node capacity depends on resolution, codec, and how many AI models you enable, so we size and validate it on your own hardware during the proof of concept. The auto-tuning installer profiles your hardware and automatically configures optimal database pools, worker counts, buffer sizes, and connection limits.
Yes. Visylix is 100% on-premise with zero cloud dependency. It is fully air-gap compatible, the entire system operates without an internet connection after initial installation. All video, AI data, face databases, and Radha AI conversations stay within your network. It supports on-premise, cloud, hybrid, and edge deployments via Docker and Kubernetes.
Visylix includes 22 production-ready AI models: face recognition (large-scale identity matching), license plate recognition (ANPR/ALPR, global formats with an India-optimized pipeline), object detection (80+ classes), person tracking (per-camera tracking plus cross-camera appearance journeys, Re-ID v1), crowd detection, PPE/safety gear detection, heat map analytics, motion detection, pose estimation (17 keypoints, fall detection), unique person counting, intrusion detection, line crossing detection (directional in/out/net counting), camera tampering detection, abandoned object detection, tailgating detection (with opt-in badge-to-body correlation), queue and wait-time analytics (with predicted wait), parking occupancy (with auto slot discovery), speed estimation, anomaly detection (semantic per-camera memory), fire and smoke detection, weapon detection, and audio event detection. Model adaptation is per camera and bounded: motion, audio, crowd, and camera-tampering detection build per-camera scene baselines, and anomaly detection trains a per-camera model from the visual embeddings of that scene. For the threshold-based detectors, Visylix surfaces retune suggestions that an operator reviews and approves. Nothing retrains itself from operator feedback, and nothing leaves your building. Detection accuracy depends on camera angle, lighting, and scene, so we validate every model on your own footage during the proof of concept.
They are the two halves of the automation. Sudarshan is the watching half: it renders live AI overlays across 12 models on the feed an operator is focused on while the rest of the video wall stays clean, it lets operators arm a watch by describing it in plain language across six condition types (faces, plates, plate patterns, objects in a zone, line crossings, and free-text semantic search), and it runs the automation rules engine, which combines AND/OR conditions and ordered temporal sequences with branching actions per condition, correlates video against access control, and can dry-run against real history before you arm it. Every alert carries a dispatch score, with an effectiveness board and one-click bounded retune, inspired by AVS-01 dispatch scoring and ISA-18.2 alarm rationalization though not a certified implementation of either. Radha is the asking half: a conversational copilot with 100+ tools that searches video by description, runs multi-step investigations, enrols faces, authors automations and briefs a shift, all in plain language. Sudarshan is included with any AI plan with no separate SKU. Physical actions notify your team; Sudarshan does not actuate hardware.
Visylix offers a free 7-day trial with 1 stream, 1 user, and 1 viewer. Paid plans start at $49/month (Starter), $99/month (Pro), and $399/month (Scale with Face Recognition AI). Enterprise plans scale up to all 22 AI analytics at custom pricing, priced by the analytics you enable. There is no per-camera licensing on any plan, so stream count never changes the price.