Workmate · ObservationPowered by Qualcomm48 TOPS INT826+ Pre-Built Behaviors3D Digital Twin

Sibox

Edge AI box transforming existing cameras into real-time operational insights.

AI Compute

48 TOPS INT8 (Dual eNPU)

Camera Streams

16+ Cameras / Box (1080p30)

Power & Temp

16 W | −20°C to 60°C

Memory & Storage

16+128 GB (SSD up to 2 TB)

Cellular Option

4G/LTE (nano-SIM slot)

Dimensions

195 × 114 × 80 mm

Install Once, Subscribe Monthly

Contact for pricing

1,000,000 – 1,500,000 VND/camera/month (>10 cams) · 24/7 VMS/Cloud & 24–48h Box Swap

Request 4-Week Pilot 26+ Behavior Catalog Watch Demo Video

01 · REALITY & INVESTMENT BARRIERS

Cameras Record 24/7, Nobody Can Watch Them All

Traditional CCTV only serves as post-incident playback, while conventional AI upgrades demand costly hardware replacements.

01 – Current Reality

4 Pain Points of Passive CCTV Monitoring

Late Detection:Incidents are usually discovered only after someone reports them; cameras are only used to rewind footage afterward.
Overloaded Operators:A single security guard cannot continuously track dozens of camera feeds throughout an entire shift.
Overwritten Evidence:By the time an investigation is needed, the video clip has often exceeded the NVR retention period.
Zero Management Metrics:Visitor counts, queue waiting times, and repeated SOP violations remain unmeasured.

02 – Traditional AI Barriers

Why Conventional Camera AI Costs Too Much

Replacing Working Cameras:Existing functional cameras are discarded for specialized AI cameras (~16,000,000 VND / camera).
Expensive GPU Servers:Face search and people analytics typically require heavy centralized GPU server infrastructure.
Perpetual Cloud Fees:Cloud video streaming and analysis charge per camera every month (~230,000 VND/cam/month) forever.
No Internet, No Alerts:When processing depends on Cloud or remote servers, any network drop blinds the entire alerting system.

02 · 05 KEY DIFFERENTIATORS

What Makes Sibox Different from Ordinary AI Boxes

Plug one compact box into your existing LAN camera network — AI analyzes video on-site and sends alerts with photo and clip evidence.

No External GPU Needed

AI runs directly on the integrated Qualcomm Dual eNPU V3 delivering up to 48 TOPS INT8 at only 16 W power consumption.

No Controller Required

Scale effortlessly by adding more Sibox nodes; the software automatically balances camera stream loads across boxes.

Works Offline When Internet Drops

Inference happens 100% on-premise; events queue locally and automatically synchronize once network connectivity returns.

Single Perception, Multiple Rules

A shared perception layer runs once per frame; add new business behavior rules purely through configuration.

Reuses Existing Cameras

Read-only RTSP stream ingestion with zero interference — your existing cameras and NVR recorders keep recording as usual.

03 · HOW SIBOX WORKS

From Video Frame to Actionable Alert in Seconds

Sends structured metadata, snapshot, and clip reference instead of raw 24/7 video — saving WAN bandwidth while preserving full evidence.

STEP 1

Camera / NVR

Uses existing RTSP cameras. NVR continues recording 24/7 without modification.

STEP 2 · SIBOX EDGE AI

5-Stage On-Device Pipeline

01 Ingest RTSP → 02 Perception & Tracking → 03 Quality Check → 04 Behavior Rules → 05 Event + Evidence

STEP 3

VMS | Mobile | Kiba

Event center, role-based access control, audit trail, and automatic task dispatch via Kiba.

STEP 4

Branch / Ward → Province IOC

Aggregates approved operational events across sites via Adaptive Integration Layer.

≤ 100 msRecognition per video frame

≤ 2 sEvent creation after confirmation

≤ 5 sEvent available on VMS / Mobile

WAN SaverSends alerts & snapshots, not raw video

04 · BUSINESS BEHAVIOR LIBRARY

21 Pre-Built Workplace Behaviors Across 05 Groups + Smart Parking, Gate LPR & 3D Digital Twin

Applicable across Public Administration, Banking Branches, F&B, Retail, Corporate Offices, and Industrial Facilities.

GROUP A

Staff Conduct & Compliance

Staff Absent from Post:Counter, desk, or gate left unattended beyond configured duration.
Unauthorized Staff Gathering:Excessive staff count congregating in a zone past time threshold.
Uniform / PPE Violation:Detects missing mandatory uniform, helmet, or safety vest.
Improper Sitting / Lying Down:Detects sitting or lying posture in restricted areas during work hours.

GROUP B

Operational SOP Compliance

Service Counter Unattended Too Long:Cashier or transaction counter without staff during operating hours.
Off-Schedule Door Open / Close:Door opened or closed outside configured time schedule.
Inspection Zone Neglected:Critical area not visited by patrol/inspection staff within required interval.

GROUP C

Customer Service Quality

Customer Waiting Too Long at Counter:Visitor waits in counter zone longer than maximum SLA time.
Queue Length Exceeded:Number of people in queue exceeds threshold for the service area.
Occupied Table Unserved:Seated guests not approached by staff within target response time.
Service Zone Overcrowded:Crowd density in lobby or service area exceeds safe capacity.

GROUP D

Business & Traffic Analytics

Zone Utilization Rate:Measures occupancy levels per zone over time to identify hot and cold areas.
Peak Hour Identification:Tracks hourly and weekly visitor peaks for staffing optimization.
Visitor In / Out Counting:Bidirectional line-crossing counts aggregated by hour and day.
Average Dwell Time:Calculates how long customers stay in a branch, store, or zone.
Table Occupancy Ratio:Real-time ratio of occupied tables vs total tables.
Movement Heatmap:Visualizes high-traffic paths and stopping points.

GROUP E

Safety & Incident Detection

Restricted Zone Intrusion:Person enters polygon-defined restricted zone during unauthorized hours.
Fall Detection:Detects person falling to the floor and remaining motionless, sending instant clip evidence.
Camera Blinded / Signal Loss:Alerts when camera view is obstructed or video feed drops.
Camera Tampering / Angle Shift:Compares against reference frame to detect camera rotation or displacement.

TRAFFIC, WATCHLIST & 3D TWINSmart Parking, Gate LPR & 3D Building TwinMisaligned / Double Parking:Detects angle deviation θ ≥ 12°, line overlap ≥ 20% held ≥ 60s (≥90% precision, ≥85% recall).
Automated Gate LPR Logging:Reads 1- and 2-line license plates (≥98% capture, ≥95% day, ≥90% night, ≤2s to VMS).
Multi-Camera Watchlist & Trace:Cross-camera timeline tracking (≥95% accuracy, ≤3s alert, non-matched faces discarded immediately).
3D Digital Twin Building Map:3D spatial model from existing cameras maintaining person IDs across floors & rooms.

IN ACTION

Watch Sibox Real-World Demo Video

Live camera feed running multiple parallel behavior rules on the on-premise Sibox management console.

05 · SOLUTION COMPARISON

Same "Camera AI" Goal — Fundamentally Different Architecture

How Sibox compares against conventional Camera AI, centralized GPU servers, and proprietary AI NVRs.

Criteria	SIBOX (Naiscorp)	Telecom Camera AI	Centralized Server AI	Proprietary AI NVR
Existing Cameras	Reused via RTSP(no replacement)	Replace with new AI cameras (~16M VND/cam)	Reused, processed on central server	Prioritizes same-brand cameras
Processing Infrastructure	48 TOPS NPU inside box, no discrete GPU	Cloud (Option 1) or GPU Server (Option 2)	Centralized GPU server cluster	NPU inside recorder
When Internet Drops	Keeps detecting locally, auto-syncs later	Depends on cloud connectivity	Depends on deployment topology	Runs locally
Business Rules	26 pre-built behaviors + no-code custom rules	People count, wait time, face recognition	Focused primarily on face recognition	Fixed vendor algorithm set
Beyond Alerts	3D Digital Twin, Kiba AI Agent, Service Robots	VMS + Cloud dashboard	Central management platform	Closed vendor ecosystem

06 · APPENDIX A2 SPECIFICATIONS

Sibox Hardware Technical Specifications

Complete hardware parameters according to official engineering specifications.

AI Accelerator	Qualcomm Dual eNPU V3; 48 TOPS INT8 / 12 TFLOPS FP16
CPU & Operating System	6-core processor / Android, Ubuntu, Linux
Memory & Expansion	16 GB RAM + 128 GB Storage; M.2 NVMe SSD (≤ 2 TB); MicroSD
Rear Interfaces & I/O	1 × RJ45 1 Gbps LAN; 4 × USB 3.0; 2 DI / 2 DO; DisplayPort; Optional 4G/LTE (nano-SIM)
Power & Environment	12 V DC, 16 W; −20°C to 60°C operating temperature; 5–95% RH
Video Decoding	16 × 1080p30 or 32 × 720p30 (H.264 / H.265 / VP9)
Dimensions (L×W×H)	195 × 114 × 80 mm
Data Security & Privacy	Read-only RTSP; local-first inference; default non-facial mode; HTTPS + mTLS + short-lived signed URLs

07 · SERVICE PACKAGES & INVESTMENT

Install Once, Subscribe to AI Monthly

Transparent two-phase investment model with 24/7 VMS/Cloud support, FPT Cloud Level-4 security storage, and 24–48h hardware replacement.

Phase 01 · One-Time (~04 Weeks)

Initial Setup & Handover

Billed once based on actual site survey and project scope — zero camera replacement required.

01. Site survey, RTSP & camera angle assessment
02. Sibox installation & existing LAN/NVR link
03. AI rule calibration, ROI zones & Cloud link
04. Operator training & technology handover

Request Site Survey

Phase 02 · Under 10 Cameras

15,000,000 VND / box / month

All-inclusive monthly AI processing subscription billed per Sibox unit for sites with under 10 cameras.

Includes 24/7 VMS & Cloud technical support
Hardware failure: new box swap within 24–48h
Event storage on FPT Cloud (InfoSec Level 4)
Optional Event Cloud: 150,000 VND/cam/month

Subscribe Now

Phase 02 · Over 10 Cameras

1 – 1.5M VND / cam / month

Flexible per-camera monthly AI subscription for larger branches, headquarters, and multi-node deployments.

1,000,000 – 1,500,000 VND / camera / month
Scale seamlessly: 01 Box (~10 cams), 02 Boxes (≤32 cams), 04 Boxes (≤64 cams)
Optional Add-ons: Kiba AI Agent & 24/7 full video cloud

Request Enterprise Quote

Your Existing Cameras Are Already Good Enough — Let Them Start Understanding!

Deploy a 4-week Sibox pilot at 01 site with zero camera replacement.