The only conversational AI copilot that runs your entire video management system — and never touches the cloud. Ask, search, investigate, automate, and administer through natural conversation.
Radha is not a chatbot bolted onto a dashboard. She commands 100+ operational tools covering every Visylix feature, answers questions from every AI analytic you have enabled, finds footage from a visual description, runs multi-step investigations with evidence timelines, sees through an on-premise vision model, listens through local speech-to-text, and keeps the site vocabulary your team teaches her. Every conversation stays inside your infrastructure.
100+
Operational Tools
22
AI Models Queryable
55+
Languages Supported
0
Cloud Dependencies
Most AI assistants answer questions. Radha executes operations, investigates incidents, writes reports, and remembers your site vocabulary. Every Visylix feature is controllable through conversation.
Talk to your VMS like you talk to a colleague. Radha understands requests like "show me all face detections from the parking lot cameras between 2pm and 4pm yesterday" and executes them instantly — in 55+ languages, replying in yours.
Radha is purpose-built on an on-premise language model that runs locally on your GPU, with optional on-premise vision and speech models beside it. Every conversation, image, and voice clip stays inside your infrastructure with zero cloud dependency — built for government, defense, banking, and any organization that requires complete data sovereignty.
Radha keeps a site vocabulary you control. Tell her once to "call this camera the gate cam" and every future conversation understands it. She remembers per-user preferences and the corrections your team gives her, and all of that memory stays on your hardware. We do not claim an accuracy improvement figure for it.
For open-ended questions Radha becomes an investigator: she plans, chains searches across sightings, detections, events and alerts, and returns findings with an evidence timeline you can replay moment by moment. The kind of capability others run in the cloud — Radha does it air-gapped.
Every important alert gets a plain-language explanation of what happened and why it fired, written by the local model seconds after the alert lands — so operators triage the alert rail without decoding raw detection data.
Radha has no delete tools, and state-changing actions never execute from model output alone. Radha proposes, names exactly what will be affected, and waits for your explicit click — the confirmation is enforced server-side with single-use tokens, so the AI cannot approve itself.
From a person's movements to a full incident investigation — here is how operators use Radha every day.
When was James recognised today?
James was recognised 4 times today: first at 08:52 at the Main Entrance, then at 09:14 in the Lobby, 13:03 at the Cafeteria camera, and last seen at 17:41 leaving through the Employee Gate. Want me to play the recording of any of these moments?
Find a person in a red jacket near the entrance this morning
I found 3 matches this morning: 08:12 at Main Entrance (strong match), 08:47 at Side Entrance A, and 11:20 at Parking Entry. The 08:12 sighting is the closest match — shall I open that recording at the exact moment?
What happened at the loading dock this afternoon? Anything unusual?
I investigated the loading dock between 12:00 and 18:00: 14 person detections (normal for a weekday), one tripwire crossing at 14:22 during a scheduled delivery, and a 6-minute loitering event at 16:48 that did not match any delivery window. I have built a timeline — the 16:48 event is worth reviewing. Play it?
Summarise my shift and email me the report
Your shift summary: 24/24 cameras online, 7 AI alerts (2 unacknowledged — both low severity), zero PPE violations, storage at 61% with 34 days of runway. Nothing needs escalation. I have generated the printable report and it is on its way to your email.
Radha does not simulate actions — she executes them against your live system, within your permissions, with confirmation gates on anything that changes state and no ability to delete.
List and play cameras, start or stop streams and recordings, capture snapshots, control playback (pause, seek, speed, fullscreen), and nudge PTZ cameras — all by asking.
"Show me the gate camera live"
"Pan camera 4 left and zoom in"
Ask about results from every AI model: face recognition, ANPR, fire and smoke, weapons, intrusion, PPE, crowd, audio events and more — with dates, cameras, and confidence.
"When was James recognised today?"
"Any fire detected overnight?"
Describe what something looks like and Radha finds it across your footage, then chains sightings, detections, events and alerts into findings plus a time-sorted evidence timeline.
"Find a man in a red jacket near the entrance"
"What happened at the dock this afternoon?"
Arm live watches and author real Sudarshan rules from a sentence, then dry-run them against history before they go live. Rules are created as safe drafts, and enabling one always requires your explicit confirmation.
"Watch for plate MH12* tonight"
"Simulate that rule against last week"
Ask why a camera is down and Radha walks the actual fault chain: reachability, credentials, stream state, recent errors, and what to try next. Then apply the fix across many cameras in one confirmed step.
"Why is the dock camera offline?"
"Restart every camera in the warehouse group"
Ask how many more cameras this server takes and Radha measures real CPU, RAM, GPU and storage headroom, then answers separately for VMS-only, single-model and multi-model AI workloads.
"How many more cameras can I add?"
"How many with face recognition on each?"
Reach recordings that live on the NVR rather than the VMS. Radha lists connected recorders, searches their archive by time, and plays a clip back inline over ONVIF replay.
"What does the NVR have for gate 2 on Tuesday?"
"Play that back from 14:30"
Send a camera to a named preset and start, stop or check a patrol tour, without hunting through the PTZ panel for the right position.
"Send the yard camera to the loading bay preset"
"Start the overnight patrol tour"
Work the alarm console conversationally: open an alarm, read its full detail, move it through triage, check sensor history, clear the notifications inbox, and play a talk-down clip to deter on site.
"What is on the alarm console right now?"
"Play the trespass warning on the yard camera"
Jump to the exact moment on any camera, bookmark recordings, search the audit log, and place tamper-evident legal holds with SHA-256 hashes — every state-changing step confirmation-gated, and nothing can be deleted.
"Show me camera 2 at 14:32 yesterday"
"Lock recording 42 as evidence"
A narrated morning briefing of cameras, alerts, PPE and storage — on demand or scheduled to your channels. One-sentence incident reports become structured, printable evidence documents.
"Summarise my shift"
"Make a report of that fire alert"
Add cameras, scan the network for ONVIF devices, create users, restrict camera access, enroll faces and plates, browse sites, groups and floor plans, test integrations, activate licences, restart services.
"Enroll James from snapshot 12"
"Let Priya see only the lobby cameras"
Radha answers questions about Visylix itself from the running system, not from invention: version, licence tier, what your entitlement includes. She remembers your site vocabulary and preferences on the box.
"What does my licence cover?"
"Call this one the gate cam from now on"
Radha was built for organizations where data privacy is not optional — and where an AI with real power needs real guardrails.
Language, vision, and speech models all run on your hardware. No API calls to external services, no telemetry, no data leaving your network — Radha works in completely air-gapped environments, from government facilities to classified networks.
Arming a rule, restarting services, running a bulk action across cameras — Radha proposes the action, names the exact target, and waits for your explicit click. The gate is enforced server-side with single-use tokens; the model cannot confirm its own actions.
Radha enforces your role-based permissions, per-camera access restrictions, and licence entitlements on every single tool call — and writes every AI action to the audit trail. Operators can never do through Radha what they could not do in the UI.
Person searches based on race, ethnicity, or religion are refused before the model ever runs. Prompt-injection attempts are blocked and audited. Radha answers only about your VMS — never the internet, never off-topic, never your system internals.
No other video management system ships a conversational AI copilot that runs entirely on-premise. Cloud vendors keep their AI in their cloud; on-premise vendors ship search boxes. Radha is a full copilot — chat, search, investigation, vision, voice, and administration — with zero cloud dependency.
Every tool executes real system commands within your permissions. When you ask Radha to enroll a person, restrict a user to specific cameras, or lock a recording as evidence, she calls the Visylix platform and confirms the result. This is not a simulation.
Radha investigates. Ask "what happened at the dock this afternoon?" and she autonomously chains searches across sightings, detections, events, and alerts into findings with a replayable evidence timeline.
Radha remembers your site. Camera nicknames, user preferences, and corrections your team gives her accumulate in on-box memory, so she uses your naming rather than the database IDs. All of that memory stays on your hardware.
Radha is governed. No delete tools at all, confirmation gates on state-changing actions, refusal of discriminatory person searches, prompt-injection blocking, licence and permission enforcement on every call, and a complete audit trail of every AI action.
Schedule a live demo and experience what it feels like to run your entire video infrastructure through conversation.
Radha is a conversational AI copilot built into the Visylix video management system. Operators control cameras, search recordings, query all 22 AI analytics models, run investigations, author automations, and administer the platform using natural language in 55+ languages. Radha runs an purpose-built on-premise language model entirely on premise, so every conversation stays within your infrastructure — it is the only VMS copilot of its kind that works fully air-gapped.
Radha commands 100+ operational tools covering every Visylix feature: live operations (streams, recordings, snapshots, playback, PTZ), AI answers across all 22 models, semantic video search, multi-step investigations, watches and Sudarshan automation rules with dry-run simulation before they go live, camera fault diagnosis and bulk actions across the fleet, server capacity and sizing questions, NVR recorder playback, PTZ presets and patrol tours, the alarm console with sensors, notifications and talk-down audio, daily briefings and incident reports, forensics and evidence locks, and full administration including camera onboarding, user management, face and plate enrollment, zones, sites and floor plans, integrations, licensing, and service restarts. She has no delete tools at all, so nothing can be removed from the system through her.
Yes. Radha performs semantic video search from plain descriptions — "find a man in a red jacket near the entrance" or "was there a white van this morning" — and returns ranked, time-stamped matches you can play instantly. The search runs on an on-premise vision-language index; no footage or query ever leaves your network.
Yes. Ask an open question like "what happened at the loading dock this afternoon?" and Radha runs an agentic multi-step investigation: it autonomously chains person sightings, AI detections, system events, alerts, and video search into findings plus a time-sorted evidence timeline, and can turn the result into a structured, printable incident report.
Yes. Radha can look at any live camera or snapshot and describe the scene using an on-premise vision model ("what is happening on the office camera right now?"), and accepts spoken questions through fully local speech-to-text. Both run on your hardware — no cloud vision or speech APIs are ever used.
Yes. Radha learns each site's vocabulary — tell her once to "call this camera the gate cam" and every future conversation understands the nickname. She remembers per-user preferences, learns from thumbs-up/down feedback and corrections, and improves the longer your team uses her. All learning is stored on your own hardware and never leaves the deployment.
No. Radha runs entirely on premise: the purpose-built on-premise language model, the vision model, and speech-to-text all execute on your hardware. Radha has no internet access by design and works in fully air-gapped environments — government facilities, defense installations, and classified networks use it without modification.
Radha enforces role-based permissions, per-camera access restrictions, and licence entitlements on every tool call, and writes every AI action to the audit trail. Radha has no delete tools at all, so nothing can be removed from the system through her. State-changing actions (arming rules, restarting services, bulk camera operations) require an explicit user confirmation that names the exact target and is enforced server-side — the model cannot approve its own actions. Person searches based on race, ethnicity, or religion are refused, and prompt-injection attempts are blocked and audited.
Cloud VMS vendors run their AI assistants in their cloud; on-premise vendors ship search boxes at most. Radha is a complete copilot — conversational control, semantic search, agentic investigation, vision, voice, reports, and administration — running 100% on premise. It answers in 55+ languages, remembers the site vocabulary your team teaches her, and is governed by confirmation gates, query moderation, and a full audit trail.