AI video intelligence · Edge-first · Nigeria first
Your cameras already saw it. Now they can tell you.
Crokta plugs in beside the CCTV you already own, runs computer vision on the appliance, and turns footage into real-time alerts with evidence and a searchable record. It keeps working when the power and the internet do not.
- No new cameras
- Raw video never leaves the site
- Outbound-only, no open ports
Works with ONVIF Profile S, RTSP, H.264, H.265, Hikvision, Dahua, Uniview, Axis, OEM rebrands, NVR channels, DVR with RTSP, PoE and Wi-Fi cameras.
The problem
Most cameras only become useful after something has gone wrong.
Organisations across West and East Africa have spent years installing CCTV. The footage is there. Nobody is watching sixteen tiles at 3 a.m., and by the time the MD asks for the clip, it takes an afternoon to find.
Nobody can watch 16 tiles for 12 hours
Attention collapses within twenty minutes. Incidents are noticed by the outcome, not the camera.
Finding the clip takes 45 minutes or more
Scrubbing NVR timelines by hand, across channels, on a screen in the guard house. Evidence arrives late or not at all.
Cloud analytics assume power and bandwidth
Grid power and 4G uplinks drop routinely. A system that stops detecting when the internet does is not security.
How it works
Sixteen cameras live and detecting in under two hours.
No rip and replace. No config files. No inbound ports. The installer scans a QR code, the box finds the cameras, you draw the zones.
- 01
Plug in beside your NVR
The Crokta Edge appliance joins your camera VLAN and discovers cameras over ONVIF, or you add RTSP streams by hand. Your NVR keeps recording. Nothing is replaced.
ONVIF discovery · RTSP · H.264/H.265 · sub-streams
- 02
Draw zones, pick a template
Polygon zones and lines on a live snapshot. Rules read like a sentence: WHEN person IN yard-perimeter DURING after-hours THEN alert Critical. Vertical templates get you 80% of the way.
16 cameras live in under 2 hours
- 03
Detect at the edge
Computer vision runs on the appliance, not in the cloud. Detection continues when the uplink drops. Every confirmed event is written to a crash-safe log before anything else happens.
≤ 300 ms frame-to-detection target
- 04
Alert with evidence, then learn
WhatsApp, push, SMS, siren or webhook. Every alert carries a thumbnail, a clip, the camera, the zone, the rule and the confidence. One tap to mark it false, and the system tightens.
≤ 8 s event-to-WhatsApp target
Degrade, don't die
Built for sites where the power and the internet do not always hold.
Detection runs on the appliance. The siren fires from the appliance. Every confirmed event is written to a crash-safe log before anything is sent anywhere. The cloud is the record of what happened, never a runtime dependency.
72 h
of event metadata buffered offline, 24 h of clips
4 h
runtime on a modest UPS, then a graceful shutdown
90 s
from boot to every stream resumed after a power cut
Zero
confirmed events lost on an ungraceful shutdown
- 18:02
Grid power drops
UPS takes over · 4 h runtime
- 18:02
Uplink lost
Ethernet → Wi-Fi → 4G · none available
- 18:41
Intrusion, gate B
Detected at edge · siren fired · event in WAL
- 18:41
Alert delivered on LAN
Guard-house console + local push
- 19:10 – 21:55
112 events buffered
Metadata 72 h · clips 24 h · nothing dropped
- 21:56
Uplink restored
Backfill ingested · one summary notification, no alert storm
Every alert carries evidence
Never a notification without a thumbnail, a clip and a reason.
Alerts land where your people already are: WhatsApp, push, SMS, email, the console, a siren, or a webhook into your own systems. Each one shows the box, the zone, the rule and the confidence, and takes one tap to acknowledge, assign, or mark false.
Under 8 seconds to WhatsApp
Design target from detection to a delivered message with the thumbnail attached.
Teach it in one tap
Mark an alert false and pick a reason. Similar events from that camera collapse for 30 minutes, and the site tightens.
Precision beats coverage
Alerts your operators will still trust in month six.
A muted system has zero value, so we measure ourselves on the false-alert rate, not the number of detectors we ship. Every site gets a 14-day tuning window and a published precision figure before contractual targets apply.
- false alerts per camera per day, from day 15
- ≤ 3
- median time from incident to evidence in hand
- < 2 min
- frame to detection on the appliance
- ≤ 300 ms
- steady-state upload per camera when nothing is happening
- < 5 kbps
Design targets from the product specification. Measured and published per site.
Detection catalogue
Start with the four every site needs. Add what your vertical demands.
Person, vehicle, intrusion, loitering and tamper on day one. PPE, man-down, fire and smoke, weapons and counting on Pro. Conservative thresholds by default; a detector that cannot hit its precision target does not ship.
Perimeter and security
The detectors every site needs on day one. Conservative thresholds, tuned per camera.
- Available at launch
Person and vehicle
Classes, tracks and dwell time for every camera.
- Available at launch
Intrusion and line-cross
Polygon zones and directional lines, per schedule.
- Available at launch
Loitering
Dwell beyond a threshold in a zone you draw.
- Available at launch
Camera tamper
Covered, defocused, repositioned or blinded.
- Pro
Weapon
Visible firearm or blade, flagged Critical.
- Roadmap
Number plates (ANPR)
Plate read at gates, matched to allow and deny lists.
Safety and compliance
For plants, depots and warehouses that need a number for the auditor every Monday.
- Pro
PPE compliance
Helmet, vest and boots by zone and shift.
- Pro
Man-down
A person on the ground who does not get up.
- Pro
Fire and smoke
Early visual detection, independent of smoke detectors.
- Pro
People and vehicle counting
Occupancy, throughput and queue length.
- Roadmap
Forklift proximity
People inside a vehicle's danger envelope.
- Roadmap
Abandoned object
A bag or box left in a zone longer than allowed.
Who it is for
Anywhere the cameras are already up and nobody is watching.
- 01
Warehouses, factories, depots
PPE, man-down, intrusion, forklift proximity
- 02
Gated estates and facility managers
Intrusion, loitering, ANPR, tailgating
- 03
Retail chains
Shrinkage cues, people counting, queues, till areas
- 04
Banks, fuel stations, telecom sites
Weapon, intrusion, tamper, ANPR
- 05
Schools, hospitals, hospitality
Intrusion, crowd, fire and smoke, falls
- 06
Local government
Fly-tipping, anti-social behaviour, ANPR
Privacy is a feature
Built to pass your IT manager, your DPO and your conscience.
Outbound-only appliance
No inbound connections, no open ports, per-device certificates, signed firmware and models.
Raw video stays on site
Only event metadata, thumbnails and short clips leave the appliance, over TLS, straight to encrypted storage.
Biometrics off by default
Face recognition is an org-level switch with a recorded consent basis, a separate store and a separate key.
Lines we will not cross
No emotion recognition, no demographic classification, no "suspicious person" scoring. Ever.
Questions
The questions Ifeoma in IT asks first.
Something else? Ask us directly.
Do I need to replace my cameras?
No. Crokta works with the IP cameras, NVRs and DVRs you already have, as long as they expose an RTSP stream or speak ONVIF. Hikvision, Dahua, Uniview, Axis and their OEM rebrands are the ones we test on every release. At onboarding, every camera gets a green, amber or red compatibility grade so you know what to expect before you pay for it.
What happens when the power or the internet goes?
Detection keeps running on the appliance. Local actions like a siren or relay still fire. Confirmed events are written to a crash-safe log and buffered for at least 72 hours, clips for 24 hours. When the uplink returns, everything is uploaded and ingested, but you get one summary notification, not a storm of stale alerts. On a modest UPS the appliance runs for four hours and shuts down gracefully on low battery.
Does my video go to the cloud?
Raw video never leaves the site. The appliance uploads event metadata, a thumbnail and a short clip for each confirmed event, and nothing when nothing is happening. Steady-state upload per camera is under 5 kbps. Live view is peer-to-peer WebRTC from the appliance to your phone or console.
How do you stop the system crying wolf?
Every new site goes through a 14-day tuning window. Days one to seven are observation mode: events are recorded, only Critical alerts are sent. Days eight to fourteen tighten thresholds using real labels from your operators. From day fifteen the contractual target applies: three or fewer false alerts per camera per day, measured and published per site.
What does the appliance need from my network?
Outbound HTTPS only. The appliance accepts no inbound connections, opens no ports, and authenticates with a per-device certificate. It fails over from Ethernet to Wi-Fi to 4G and is aware of data caps.
What about face recognition?
Off by default, at the organisation level. Turning it on records the legal basis, stores biometric data in a separate store under a separate key, and audits every access. We do not do emotion recognition, demographic classification or any kind of "suspicious person" scoring, and we will not build them.
Who installs it?
Usually your existing CCTV installer, who we train and certify. The install flow is designed to take under two hours for sixteen cameras: scan a QR code to claim the box, review discovered cameras, pick a template, draw zones, start observation mode, hand over.
Pilot programme
Make the cameras you already own earn their keep.
Start with a 14-day paid pilot on one site. We tune it with your operators, publish the false-alert rate honestly, and you decide from day fifteen.