The AI video management system built for plants, not data centres.
Your cameras already see everything. Nobody is watching. Neurabit VMS records every camera, watches all of them continuously, decides what is genuinely out of place, and tells the right person within seconds — on hardware inside your own building.
- Works with the internet down
- Footage never leaves your site
- DPIIT recognised
- Deployed in 7+ countries
16 of 32 channels · analysed continuously
Footage gets reviewed after something happens. Never before.
of footage a 32-camera site produces every single day
is what one supervisor can meaningfully watch at once
before attention measurably degrades
Why most AI CCTV analytics fail in factories
Most plants that have tried this have the same story. A vendor installed something, it worked in the demo, and within six weeks nobody opened the alerts. Three reasons, and it is the same three every time.
It cried wolf.
Forty false alerts per camera per day trains a supervisor to ignore the system. Once that trust is gone, no later improvement recovers it.
We target under one false positive per camera per day on critical alerts, and every rule shows you its own false-positive rate.
It needed the internet.
32 cameras at recording quality is roughly 128 Mbps of sustained upload. No ordinary Indian leased line carries that alongside everything else the site needs, and when the link drops, monitoring stops.
Everything runs on hardware inside your own building. Full function with the WAN down — alerts queue and deliver on reconnect.
It could not be changed.
Twelve fixed use cases. Need a thirteenth, or want the loitering threshold moved from five minutes to fifteen? That is a change request and a quotation.
Your team writes and tunes the rules in plain language, and can upload your own ONNX models. No ticket, no waiting on our roadmap.
We built this because we kept hitting all three while deploying computer vision for industrial clients.
17 use cases, live from day one
Perimeter & Access
- Perimeter intrusion and fence climbing during non-working hours
- Unauthorised entry to restricted areas: server room, cash room, stores, electrical rooms
- Unauthorised vehicle entry, exit or movement outside approved timings
- Visitors in unauthorised areas or overstaying their approved duration
- Staff leaving without completing an exit punch
Workforce Compliance
- Staff on premises after hours without authorisation
- Loitering in canteens and break areas beyond designated break timings
- Gathering in corridors, staircases, parking or wash areas during working hours
- Prolonged inactivity at workstations
- Security personnel absent from post or missing patrol rounds
Safety
- Slips, falls and collapses requiring immediate response
- Physical altercations and workplace misconduct
- Fire and smoke, including on holidays and after hours
- PPE compliance: helmet, vest, gloves, boots, harness
System Integrity
- Camera tampering, obstruction, defocus or feed loss
- Loading and unloading outside approved timings
- Unattended bags, cartons or objects in common areas
Every rule is schedule-aware. The same person in the canteen is fine at 1pm and a flag at 4pm, because the system knows your shifts, break windows, holiday calendar and who is authorised where.
A plant is the hardest case, not the only one
Anywhere with cameras, shifts and a boundary is the same problem. We lead with manufacturing because dust, glare and contractors in street clothes are the hardest conditions to hold accuracy in — an office lobby is a far easier room to get right.
Manufacturing plants
Dust, shift patterns and contractors in street clothes. PPE and near-miss evidence carry the audit.
- PPE compliance
- Slips & falls
- Fire & smoke
- Restricted-area entry
- Crowding & gathering
- Loading off-hours
- Prolonged inactivity
- Camera tampering
Warehouses & logistics
Bay discipline and vehicle timings decide throughput; most losses happen outside approved hours.
- Loading off-hours
- Vehicle movement off-hours
- PPE compliance
- Slips & falls
- Unattended objects
- After-hours presence
- Camera tampering
- Restricted-area entry
Corporate offices & IT parks
Access control and visitor discipline matter far more than PPE. Server and records rooms are the sensitive zones.
- After-hours presence
- Restricted-area entry
- Visitor overstay
- Exit without punch
- Unattended objects
- Crowding & gathering
- Fire & smoke
- Camera tampering
Hospitals & healthcare
Patient falls and altercations need a response in seconds, and restricted wings need a real boundary.
- Slips & falls
- Altercations
- Restricted-area entry
- Visitor overstay
- After-hours presence
- Fire & smoke
- Unattended objects
- Guard off post
Schools & campuses
Perimeter and after-hours presence first, then anything involving a crowd forming where it should not.
- Perimeter intrusion
- Restricted-area entry
- Altercations
- Crowding & gathering
- Visitor overstay
- After-hours presence
- Fire & smoke
- Unattended objects
Retail & showrooms
Shrinkage happens after hours and at the stockroom door; crowding and abandoned bags are the floor risks.
- After-hours presence
- Restricted-area entry
- Unattended objects
- Altercations
- Loitering past a limit
- Crowding & gathering
- Camera tampering
- Exit without punch
Hotels & hospitality
Back-of-house access, guest-area incidents, and whether the person on the desk is actually on the desk.
- Restricted-area entry
- Visitor overstay
- Guard off post
- Altercations
- Unattended objects
- Fire & smoke
- After-hours presence
- Exit without punch
Gated communities & residential
Boundary, gate and guard discipline — the three things residents actually pay the association for.
- Perimeter intrusion
- Vehicle movement off-hours
- Visitor overstay
- Guard off post
- Unattended objects
- Fire & smoke
- Camera tampering
- Altercations
All 17 scenarios are available in every deployment. The tags above are the ones that earn their place in each kind of building — what changes between them is the zones you draw, the schedules you enter, and which rules you actually arm.
Record. Search. Respond. Prove.
A video management system, not just analytics
Analytics without an archive is half a product. When something happens, you need the footage.
Recording that fits the camera
Continuous, motion, event or scheduled recording per camera, with retention set by days or by disk quota, whichever binds first. Storage forecasting warns you before retention drops below policy, not after footage is already gone.
A timeline that tells the truth
Scrub from thirty days down to one second. Synchronised playback across cameras on a single clock. Gaps where a camera was offline are shown explicitly, never silently skipped, because an investigator needs to know footage is missing.
Evidence that holds up
Export any clip with burned-in timestamp and camera ID, a SHA-256 manifest and a chain-of-custody statement naming who exported what and when. Every view and every export is written to an append-only audit log.
An archive that survives a power cut
The segment index is crash-safe and rebuildable. We test this by killing the recorder at fifty random points and verifying the archive is readable every time.
Find the person, not the timestamp
Reviewing footage by scrubbing is the slowest part of any investigation.
Attribute search
Filter across every camera and every day by what someone looked like: upper and lower clothing colour, whether they carried a bag, helmet, vest. Vehicle type and colour. Results come back as a gallery you can jump into the timeline from.
Natural language
Ask in plain English, on the web or over WhatsApp. How many PPE violations in the dye house last week? Show me vehicles at the loading bay after 10pm on Tuesday. The system answers from your own records and shows you which ones it used.
Cross-camera tracking
Follow a person or vehicle across the site as a single timeline rather than a dozen separate clips.
An alert is the start, not the end
Most systems stop at notification. That is where the useful part begins.
Route it
WhatsApp, SMS, email, mobile push, control-room queue, PA talk-down, or a dry-contact relay to a siren or gate.
Escalate it
No acknowledgement inside the SLA and it goes to the shift supervisor, then the plant or facility head, with every step recorded.
Close it
Acknowledge and dispose from inside the WhatsApp thread. Every closure needs a reason code, so you end up with data rather than a feeling.
Build the workflow yourself
Trigger, conditions, enrichment, routing, escalation, closure, on a visual canvas. Connect it to ServiceNow, Jira, Teams, Google Sheets or your own systems.
The report your auditor accepts
Safety and security heads do not buy video software. They buy evidence.
A monthly compliance report
PPE compliance by zone, shift and contractor, near-miss counts trending over time, and an incident register where every line links to the video.
Leading indicators, not just incidents
Near-miss frequency, response times, which zones and which shifts drive your numbers.
MTTA and MTTR per rule, per zone, per shift
You find out which alerts are being answered and which are being ignored.
No change request. No waiting on our roadmap.
This is what most vendors do not offer, and it is the reason we built the product this way.
Your team writes the rules
Pick a scenario, choose cameras and zones, set the schedule, set the threshold in plain language. Not a config file. Not a support ticket.
Your team tunes it
Every rule shows its false-positive rate. Change a threshold and see immediately what the last seven days would have produced at the new setting, before you commit.
Your team can ask why
Every alert carries its reasoning: which rule, which detections, which conditions passed. And the reverse, which matters more: when something should have alerted and did not, the system names the condition that blocked it. Usually a schedule or a zone boundary, not the AI.
Bring your own models
The pipeline accepts standard ONNX models. Upload one, the system builds and optimises it for your hardware, versions it, and lets you roll back. Chain models together visually: detect, crop, classify, count.
Test before you arm
Run any new rule or model against recorded footage and see exactly what it would have caught, before it goes live.
Replayed against the last 7 days of recorded footage before it goes live — 3 matches, 0 false alarms.
Illustrative — this is the rule editor your own team uses.
Why on-premise wins
We build on NVIDIA Jetson and on-premise x86 with NVIDIA GPUs. Everything runs at your site.
Internet requirement
Upload, 32 cameras
When the link drops
Where footage lives
Data residency, DPDP
Cost model
The architectural reason it works
We do not run every model on every frame of every camera. A lightweight detector runs continuously across all channels; heavier models run only on cropped regions, and only when a condition is met. That cascade is what makes 24 to 32 analysed channels viable on a single edge device, and what keeps per-camera cost sane.
One plant or fifty
Multi-site from one console
Site health, alert volume and compliance trends across your estate, with each site still fully autonomous.
Configuration as code
Commission site one properly, then replicate its zones, schedules, rules and templates across the rest in an afternoon instead of a week.
Updates that roll back themselves
Staged rollout, health-checked, automatic rollback on any site that fails.
Integrations
HRMS punch reconciliation, visitor management, gate pass, access control. Badge or facial identity binding, consent-managed under the DPDP Act, with templates that never leave your site.
What it actually looks like
Nobody goes live on day one. Be suspicious of anyone who says otherwise.
- Install, connect cameras, verify streamsHalf a day
- Draw zones, enter shifts and holidays1–2 days
- Switch on scenario templatesHalf a day
- Recording and alertingDay 3
- Tuning against real site conditions2–4 weeks
- Alerts your team trusts and acts onWeek 4–6
The gap between day 3 and week 5 decides whether a deployment succeeds. On day 3 the system works but flags shadows as people and cartons as abandoned bags. Weeks one to four are spent marking false positives and tuning until it settles. We stay with you through it. That is not a support add-on, it is how the product is designed to be deployed.
Three things that decide the outcome
Camera placement matters more than the software
A camera pointed into a bright doorway, or mounted too high to resolve a person, will underperform regardless of the model. We run a site survey before quoting and will tell you which of your cameras need moving.
Someone has to own the alerts
A system nobody is responsible for answering becomes wallpaper in six weeks.
We need footage from your site early
Models tune to your dust, your lighting, your uniforms, your contractors in street clothes. Every site differs, and the first two weeks are how the system adapts to yours.
Being clear about the boundary saves everyone time.
It does not read job cards, verify chemical dosing sequences, or check that a pH reading was logged. Those are ERP and checklist problems, not camera problems.
It does not fix bad cameras.
How this differs from Milestone, Genetec or Verkada
Those are strong products, built for enterprise campuses and cloud deployment. We are built for Indian industrial sites.
Full function without internet
Not a degraded offline mode. Recording, detection, rules and local alerting all continue with the WAN down.
Two ways to buy it
Own the edge hardware outright, or take it on a per-camera subscription with the hardware included. Either way it sits on your premises.
Rules your own team changes
Thresholds, schedules, zones and even the AI models themselves — no change request, no quotation.
Priced for an Indian cost base
Quoted against your actual camera count after a survey — not a global list price converted into rupees.
Two ways to buy it
Which one is right for you is a budget question, not a product question — the system is identical either way. Pick the one that matches how your plant spends.
Own it
One-time hardware capex, plus an annual licence per camera.
Best for: Sites with a capex budget and a long horizon.
- The edge hardware is bought once and is yours outright — an asset on your books, not a rental.
- Software is licensed annually per camera, so the cost tracks the cameras you actually analyse.
- The licence covers updates, model improvements and support, through the tuning weeks and after.
- Lowest total cost of ownership from roughly year three onward.
Add cameras later at the same per-camera licence rate.
Subscribe
Per camera, per month. Hardware included.
Best for: Opex-only budgets, leased premises, or a first site ahead of a wider rollout.
- No capital outlay. The edge hardware is supplied, maintained and replaced by us as part of the subscription.
- One predictable per-camera monthly figure covering hardware, software, updates and support.
- Scale cameras up or down as the site changes, instead of re-quoting a hardware purchase.
- The fastest route to a live site when procurement is the bottleneck.
The hardware still sits on your premises throughout — and so does your footage.
- Footage is recorded and processed on your premises on either tier. Neither one sends video to us.
- Both are quoted after a site survey, because camera count, retention and which analytics you need all move the number.
- No per-camera cloud fee, no bandwidth bill, and no charge for storage you already own.
For the person who will evaluate it
- Cameras
- Any ONVIF-compliant IP camera, RTSP
- Compute
- NVIDIA Jetson Orin, or x86 with NVIDIA GPU
- Analysed channels per device
- Typically 24–32 on mid-range Jetson; more on x86 with discrete GPU
- Recorded channels
- Higher than analysed; disk-bound rather than compute-bound
- Recording
- Continuous, motion, event or scheduled, per camera
- Retention
- By days or disk quota, with forecasting
- Deployment
- On-premise. Optional cloud console for multi-site
- Offline operation
- Full function with the WAN down
- Custom models
- Standard ONNX, uploaded through the interface
- Access control
- Role-based, per-camera permissions, SSO via OIDC and SAML
- Audit
- Append-only, hash-chained
- Languages
- English, Hindi, and regional
Neurabit is a DPIIT-recognised deep-tech company headquartered in Sambalpur, Odisha, with production AI systems deployed across 7+ countries. We build edge AI, computer vision and physical AI systems. This product exists because we kept rebuilding the same pieces for industrial clients and decided to build them properly, once.
AI that works where the internet doesn't.
Get a site survey and a sized quote.
We do not quote from a price list. Tell us what your site looks like and we will size the hardware, tell you which cameras are usable as they are, and walk you through what the first six weeks would actually involve.
- We tell you which of your existing cameras need moving — before you spend anything.
- Two ways to buy: own the hardware outright, or subscribe per camera with it included.
- A real engineer replies within one business day, not a sales sequence.
Frequently asked questions
A VMS records and manages footage from your cameras. An AI video management system adds continuous analysis on top: it watches every stream, recognises defined situations, and raises an alert with the evidence attached, instead of waiting for someone to review footage afterwards.
Yes. Everything runs on a device at your site. Recording, detection, rules and local alerting continue with the WAN down. Notifications queue and deliver in order once connectivity returns.
Any ONVIF-compliant IP camera with an RTSP stream, which covers most commercial and industrial CCTV installed in the last decade. We verify each camera during the site survey, because resolution, mounting angle and keyframe settings all affect what is achievable.
It depends on hardware and how many analytics run per camera. A mid-range NVIDIA Jetson typically handles 24 to 32 analysed channels, and can record more than it analyses. We size against your actual camera count and use cases rather than quoting a generic number.
There are two ways to buy it. Either a one-time hardware capex that you own outright plus an annual software licence per camera, or a single per-camera per-month subscription with the hardware included and no capital outlay. We quote either after a site survey, since camera count, retention and which analytics you need all move the number.
Almost always false positive rates, not detection accuracy. A system at high accuracy that still fires forty false alerts per camera per day gets ignored within a fortnight. We target under one false positive per camera per day for critical alerts, and the tuning period exists to get there.
No. Video is recorded and processed on-site. Only alert notifications leave, and only to recipients you configure.
Yes. The pipeline accepts standard ONNX models with a manifest describing inputs and outputs. Upload it, and the system builds and optimises it for your hardware.
Those are strong products built for enterprise campuses and cloud deployment. We are built for Indian industrial sites: full function without internet, hardware that sits on your premises whether you buy it or subscribe to it, rules and models your own team can change, and a price structured for an Indian cost base.
Still have a question we have not answered? Ask it in the survey form and an engineer will answer it directly.
Request a site survey →17 scenarios, running offline, inside your building. We quote after a site survey.
Size it for your site
Request a site survey →