Products / SENTINEL
S.A.L.T CI SENTINEL
Intelligence Surveillance System. Turns the cameras you already own into a sovereign eye that detects, alerts, searches and writes its own reports, on your own servers, with no camera replacement and no video leaving the premises.
Cameras that record everything and understand it
Most CCTV networks record everything and understand nothing. Footage is reviewed only after something has gone wrong, no team can watch hundreds of screens, and finding one event in a week of video can take days. SENTINEL adds a real-time and investigative layer on top of the camera network an organisation already owns.
detector families, switched on per camera, zone and schedule
behaviours, from a fall or a tailgate to grass that needs cutting
operating modes for duty officers, supervisors and analysts
alert format for every detector: one console, one search, one report engine
languages for every report and screen: Bahasa Melayu and English
video copies sent outside the site; nothing goes to a cloud unless you allow it
No camera replacement
IP cameras over RTSP and ONVIF, analogue cameras through encoders, and any VMS or NVR. Your VMS stays the system of record.
Modular detection
Twelve detector families, switched on per camera, zone and schedule. New detectors drop in without changing the platform.
One alert, one console
Every detector produces the same alert, so one console, one rule engine, one search and one report engine serve every camera.
Reports, not raw detections
Incident narratives, shift summaries and investigation reports written by AI, in Bahasa Melayu or English.
Investigations in minutes
Search footage by object, face, plate, colour, place and time, or simply ask in plain language.
Sovereign and lawful by design
On premises or air-gapped, with roles, retention, privacy masking and a full audit trail, aligned with Malaysia's PDPA.
People stay in charge
SENTINEL detects, ranks and explains. A person confirms before any action is taken against anyone, and every face or plate match shows its evidence and confidence. From one building to a city: a single server, edge boxes beside camera clusters, or many sites linked through TERAS federation.
Most camera networks are recorders, not sensors
Organisations have invested heavily in cameras. What they rarely have is a way to turn that video into timely action, evidence and insight without buying everything again.
Screens nobody watches
One operator can follow a handful of feeds. Hundreds of cameras are, in practice, unwatched until someone asks for footage.
Investigations that take days
Finding a white van, a person or a moment across many cameras means scrubbing timelines by hand, camera by camera.
Analytics that lock you in
New cameras, one brand, a vendor gateway, per-camera monthly fees, and a roadmap you cannot influence.
Data that leaves your control
Cloud analytics send faces, plates and video to servers abroad, a real concern for government, defence and regulated sectors.
How the day changes
| Moment | Today, with cameras only | With SENTINEL |
|---|---|---|
| Something happens | Seen only if an operator happens to be looking | Detected in seconds, ranked by severity, pushed to the right person |
| An alert fires | A beep and a camera number | An evidence frame, a short clip, an AI summary and three clear actions |
| Nobody responds | The alert is lost in the noise | Escalation: operator, then supervisor, then duty officer, until acknowledged |
| A complaint arrives | Hours or days of manual scrubbing | Plain-language search across cameras, a cross-camera timeline in minutes |
| Management asks what happened | Someone writes it up from memory | An AI incident report on your template, in Malay or English, with evidence hashes |
| The grass, the leak, the broken fence | Found when someone complains | Detected by the same cameras and sent as a work order |
The SENTINEL approach
Keep the cameras, keep the recorders, keep the VMS. Add intelligence on top, on your own servers, and let every detector speak the same language.
A pipeline from pixels to decisions
Video comes in, the detector registry runs the right detectors on it, every result becomes one standard alert, and the three operating modes put those alerts to work. Reasoning (confirmations, summaries, reports and plain-language search) comes from TERAS, S.A.L.T CI's sovereign AI core.
Live streams, recorded footage, uploads
12 families, 50+ behaviours
Zones, schedules, rules, watchlists
Console, phone, WhatsApp, Teams, radio
Search, timelines, cases, evidence
AI narratives, briefings, trends
| Component | Responsibility |
|---|---|
| Ingestion | Connects to cameras over RTSP and ONVIF, to VMS and NVR APIs and to uploaded files; samples frames per detector; reconnects automatically and reports camera health |
| Detector registry | Loads the detectors enabled in the site's deployment profile, per camera, zone and schedule; a new detector is a module dropped into the registry |
| Detection engine | Runs detection, recognition, segmentation and tracking models on GPU or CPU; applies zones, line crossings and dwell times; tracks objects across frames and cameras |
| Rule engine | Combines detections into events, applies thresholds, severities and quiet hours, and starts escalation |
| Alert, evidence and search | Stores every alert with its evidence, deduplicates repeats, delivers to people and systems, and indexes attributes and visual features for search |
| Report engine | Asks TERAS models to write incident narratives, shift summaries and investigation reports on the organisation's templates |
| Console, API and MCP | Watch, Monitor and Intel in the browser, on video walls and on phones; every capability as REST endpoints and MCP tools |
Eight design principles
Use what is installed
Cameras, recorders and VMS stay. SENTINEL never forces a camera brand.
Detectors are modules
A site runs only those it needs, where they make sense.
One alert format
From a plate read to a fall, every detection is the same structured alert.
Video stays where it is
Live analysis reads streams; investigations fetch only the footage they need.
Sovereign by default
Video, faces, plates and alerts are processed and stored on your servers.
People decide
SENTINEL detects, ranks and explains; a person confirms before action is taken.
Lawful and proportionate
Each detector has a stated purpose, retention and access rules.
Degrade, never break
If a camera, detector or GPU fails, the rest keep running and the failure shows.
Twelve families, fifty-plus behaviours
Detectors are grouped in families. Each behaviour is enabled per camera, zone and schedule with its own thresholds. A site runs only what it needs.
| Family | Behaviours | Typical places |
|---|---|---|
| Person | Presence and counting, zone intrusion, line crossing, loitering, running, falls, tailgating, crowding and density, wrong-way movement | Entrances, perimeters, corridors, halls, wards |
| Face | Detection, recognition against enrolled people, watchlist matches, unknown-person alerts, privacy blurring of everyone else | Entrances, lobbies, checkpoints |
| Licence plate (ANPR) | Plate reading in Malaysian formats, allow lists and watchlists, entry and exit logs, dwell time | Gates, car parks, roads |
| Vehicle | Type and colour, counting and flow, illegal parking, wrong way, stopped vehicles, speed estimate | Roads, junctions, car parks, loading bays |
| Object | Abandoned or removed objects, suspicious items, waste and illegal dumping, overflowing bins, blocked exits | Lobbies, public areas, roadsides |
| Animal | Intrusion and damage by monkeys, stray dogs, wild boar and snakes; counts by species | Campuses, housing, plantations |
| Vegetation | Grass and hedge height, overgrowth onto walkways, fallen trees and branches | Grounds, roadsides, parks |
| Structural | Cracks, damage, leaks and flooding, fallen ceilings, broken fences and barriers | Buildings, drains, perimeters |
| Fire and smoke | Flames and smoke indoors and outdoors, open burning, from camera video alone | Labs, kitchens, stores, open land |
| Education | Exam-hall integrity (phones, unusual movement, seat changes), attendance counts, room occupancy | Exam halls, lecture rooms |
| Safety and PPE | Helmets, vests and other protective equipment, people near moving machines, restricted-area entry | Factories, sites, plants |
| Camera health | Camera offline, blocked, moved, out of focus or tampered with | Every camera |
Attendance from the same cameras
Once people are enrolled with consent, face recognition at chosen entry points logs arrival, departure and time on site, replacing manual registers and stopping proxy attendance, with no new hardware.
Built for local conditions
Malaysian plate formats, wildlife that actually visits Malaysian sites, monsoon flooding, grounds that grow fast, and reports in Bahasa Melayu as well as English.
How detection actually works
Detection, recognition, plate-reading and segmentation on every sampled frame
A vision-language model through TERAS confirms frames that need understanding
Dwell, loitering, direction and abandonment measured; re-identification across cameras
Thresholds, zones and schedules per camera; fine-tuning on the site's own footage
A vision check that cuts false alarms
Detector: possible flame, confidence 0.41. Vision check via TERAS: "Sun glare on a windscreen. No flame or smoke visible." Result: alert suppressed, recorded as a rejected detection for tuning.
The duty officer's console: see it, act on it
Large, clear alert cards. Three actions only: view footage, acknowledge, escalate. Bilingual in Bahasa Melayu and English, on any screen from a control-room wall to a tablet, with a PIN gate in front of deeper analysis.
Detections on the picture
Boxes, track IDs, zones and lines drawn live, so the operator sees why an alert fired, not just that it did.
A feed that ranks itself
High first. Every card carries an evidence frame, a short clip, the confidence and an AI summary in plain words.
One click, recorded
Acknowledge, assign, escalate or mark as a false alarm. Every action is logged, and false alarms tune the thresholds.
Bahasa Melayu · English
One toggle switches every label, alert and summary, so each officer works in the language they think in.
PIN-gated depth
Duty officers stay in Watch; moving into Monitor or Intel needs a PIN or the right role, and the switch is audited.
Alerts that reach people
A patrol app with evidence and route, and the same alert in WhatsApp, Teams, Telegram, e-mail or radio dispatch.
For supervisors: the whole site by domain
Monitor mode organises everything SENTINEL sees into domain dashboards: Live Operations, Security, Identity and Access, Vehicle and Parking, Safety and Risk, Facilities, Education and Reports. Security, facilities, safety and management each get their own view of the same cameras.
Rules people actually write
| Rule | Detectors combined | Typical action |
|---|---|---|
| Vehicle not on the allow list at the gate after 23:00 | plate.unknown + time window + camera | Gate officer, then supervisor |
| Same plate entering twice without leaving | plate.sequence across entry and exit cameras | Flag for investigation |
| Person in the substation zone outside working hours | person.intrusion + zone + schedule + allow list | High alert, lock door, radio |
| Person near a moving forklift without a vest | ppe.vest + vehicle.forklift + proximity | Supervisor and siren |
| Fall in a ward, confirmed by vision check | person.fall + vlm.confirm | Nurse station and phone |
Shift and daily summaries
Handover notes that write themselves, with the incidents that need follow-up at the top.
Weekly AI briefings
Trends, hotspots, peak hours, parking occupancy and recurring issues for management.
Bahasa Melayu or English
Every report in either language, on the organisation's template, with evidence frames and confidence.
For analysts: the investigation workspace
A SOC dashboard across all cameras, an analysis studio for recorded and uploaded footage, and AI intelligence reports for every run. Most incidents are investigated after the fact; Intel mode makes answering "what happened?" take minutes, without copying a week of video anywhere.
A question in plain words, or a snapshot
Matches across cameras and time
Re-identify the subject camera to camera
A case timeline with notes and evidence
An AI incident report on your template
Clips and frames with SHA-256 hashes
One product, configured for every sector
Any organisation that runs a camera network and needs it to do more than record. The detectors, rules and reports differ by site; the product does not.
Universities and schools
Campus safety, gates and parking, hostels, exam halls, wildlife, grounds upkeep, attendance.
Local authorities
Public CCTV, traffic and junctions, parking enforcement, markets, illegal dumping, flood points.
Police, defence and security
Perimeters, checkpoints, watchlists, cross-camera investigations, air-gapped command rooms.
Factories and plants
Safety zones, PPE, forklifts near people, fire and smoke, restricted areas, loading bays.
Critical infrastructure
Substations, water plants, ports and airports; intrusion and camera tampering.
Commercial and retail
Access, occupancy and queues, after-hours intrusion, incident evidence for claims.
Hospitals and care
Falls, wandering, restricted wards, car parks, visitor flow.
Estates and townships
Guardhouses, visitor plates, perimeter fences, common areas, stray animals.
A plant at 2 am
A person enters the press-line zone without a helmet while the line runs. SENTINEL raises a high alert, the supervisor's phone rings, and the morning report records it with the clip.
A council junction
A lorry stops on a yellow box during rush hour and rubbish is left by the roadside. Traffic sees the first; TERUSAN opens a work order for the second.
A campus car park
A plate on the watchlist enters at 22:14. The gate officer is told, the vehicle is followed across three cameras, and an evidence pack is sealed for the police.
Fits the systems you already run
SENTINEL reads from what is there, writes to where people already work, and exposes every capability as REST endpoints and MCP tools so other applications and AI agents can use it.
| System | How SENTINEL works with it |
|---|---|
| Cameras | Any IP camera with RTSP; ONVIF Profile S and T for discovery and control; analogue cameras through encoders; body-worn, drone and mobile feeds |
| VMS and NVR | Direct RTSP, VMS re-streams, or the VMS's own API for recorded footage, without replacing them |
| TERAS | AI models for summaries, reports and search; vision-language checks; single sign-on, roles and audit; agents that call SENTINEL tools |
| PERISAI SOC Hub | Physical alerts beside cyber alerts for one security picture, with correlated incidents |
| TERUSAN | Alerts start workflows: a fallen tree becomes a grounds work order, a broken fence a maintenance ticket |
| Access control and barriers | Open a barrier for an allowed plate, lock a door on an intrusion, or log a mismatch between a badge and a face |
| Monitoring and other systems | Camera, detector and server health to the monitoring stack; signed webhooks and REST for ticketing and command-centre systems |
Powerful enough to matter, governed enough to trust
Surveillance touches people's rights. SENTINEL is designed so that an organisation can use it lawfully and proportionately, and show that it does.
Purpose per detector
Each detector is enabled with a stated purpose and scope; the purpose is shown on every alert it raises.
Data stays on site
Video, faces, plates and alerts are processed and stored on the organisation's servers; air-gapped operation is supported.
Least access
Roles decide who may see live video, faces, plates, search results and exports; sensitive views can require a reason.
Privacy masking
Faces of people not on a watchlist, windows of private property and screens can be blurred in live view and exports.
Retention
Separate retention periods for alerts, snapshots, clips, enrolments and cases, with automatic deletion.
Human decision
Face and plate matches show the evidence and score; acting on a match requires a person's confirmation.
Accountability
Every view, search, export and change is recorded in a tamper-evident audit log.
Enrolment control
Faces and plates are enrolled only by authorised roles, with a reason and an expiry; individuals can be removed on request.
Accuracy checks
Detector accuracy is measured on the site's own footage before go-live and reviewed regularly, including false-match rates.
PDPA-aligned by design
SENTINEL supports compliance with Malaysia's Personal Data Protection Act and the organisation's own security and privacy policies. The organisation remains responsible for how it uses the system; S.A.L.T CI provides a data-protection checklist for every deployment.
Proven on your footage before it goes live
SENTINEL is not tied to any camera brand, server or GPU. It runs on bare metal, virtual machines, private or public cloud, or Kubernetes, and every deployment is measured on the site's own video first.
| Scale | Cameras | Typical hardware |
|---|---|---|
| Small | Up to 50 | One server with one GPU, or CPU only for light detectors |
| Medium | 50–500 | A few GPU servers or edge boxes, plus TERAS for reporting and search |
| Large | 500+ | GPU clusters or edge fleets across sites, with central search and command view through TERAS federation |
Edge
A small GPU box beside a group of cameras, sending only alerts.
Single site
One server, or a few, running SENTINEL and TERAS.
Multi-site
Sites linked through TERAS federation to a central command view.
Hybrid
Edge analysis with central search, reports and dashboards.
Fully offline
Air-gapped sites; updates by signed offline bundle.
Camera inventory, stream checks and accuracy on the site's own video
Containers beside TERAS; one script detects GPUs
Discover ONVIF, add RTSP or import from the VMS
Detectors, zones, schedules, thresholds, watchlists
Alerts reviewed, not escalated; thresholds tuned
Escalation on; Watch, Monitor and Intel users trained
An honest note on accuracy
Any AI surveillance system starts noisy on a cold deployment. That is why SENTINEL measures accuracy on the site's own footage, tunes thresholds per camera and zone, enrols the allow lists, and agrees an accuracy target per detector before go-live, then keeps measuring.
Ways to buy
A site licence the organisation owns, sized by cameras and detector families, with annual support and updates. Delivered through authorised S.A.L.T CI partners. No per-camera cloud fees.
What other approaches leave out
SENTINEL compared with the common ways organisations add intelligence to CCTV. Categories describe typical offerings; individual products differ.
| Capability | Cloud video analytics | Analytics in the VMS | AI cameras | S.A.L.T SENTINEL |
|---|---|---|---|---|
| Works with existing cameras of any brand | Often needs a gateway | Within that VMS | New cameras | Yes |
| Video and detections stay on site | No | Yes | Varies | Yes |
| Runs fully air-gapped | No | Varies | Varies | Yes |
| One alert format and console for all detectors | Varies | Varies | No | Yes |
| AI-written incident reports in Malay and English | Some | No | No | Yes |
| Plain-language search of footage | Some | Some | No | Yes |
| Vision-language checks to cut false alarms | Some | No | No | Yes |
| Workflows, work orders, cyber-physical correlation | Integrations | Some | No | Yes: TERUSAN, PERISAI |
| Add new detectors for local needs | Vendor roadmap | Some | No | Yes: detector kit |
| No per-camera cloud subscription | No | Varies | Yes | Yes |
The cameras are already there. SENTINEL makes them see.
Start with a site assessment on footage from your own cameras, pilot a handful of cameras in shadow mode, then switch on escalation and grow camera by camera, site by site.
Part of one sovereign platform
SENTINEL runs on the TERAS core with the same sign-in, audit and models as every other S.A.L.T CI application.