Skip to content

Solution Summary

Project Overview

Project: ABV-SECOPS-AI-2025 - PHASE 4: AI Edge Networking Architecture Scope: Mumbai Hub + Chennai Hub edge AI deployment with observability fusion

This appendix consolidates the technical solution across all use cases and the edge AI platform architecture for quick reference.

Solution Deliverables

Chapter 2: Use Case Architecture

Section 2.1: UC1 - Intelligent Physical Security

2.1.1 Security Functions & AI Models - 6 security functions: Perimeter intrusion, loitering, tailgating, LPR, crowd density, access control - AI models: YOLO v8n (person detection), DeepSORT (tracking), PPE CNN, LPR pipeline - Performance targets: <500ms response, <5% false positive rate - Hardware: NVIDIA L4 24GB GPU on UCS XE130c M8

2.1.2 Camera Deployment & Network Architecture - Complete UCS XE9305 + UCS XE130c M8 specifications - 135 cameras per site (60 indoor, 40 outdoor PTZ, 20 4K LPR, 10 thermal) - 5-layer network architecture (Camera → Cat9300 → Cat9500 → Edge AI → Observability) - 4ms camera-to-inference latency (IDF Room Floor 3 co-location) - Complete Bill of Materials with vendor-neutral pricing

2.1.3 Observability Integration - Complete 500ms detection → response timeline - Multi-source validation: ISE pxGrid (50ms), Splunk MLTK (100ms), ThousandEyes (80ms), AppDynamics (90ms) - Complete JSON API payloads for all integrations - Automated actions: FTD blocking, XDR incidents, ServiceNow tickets, Webex alerts - Real-world scenario walkthrough: Loading dock perimeter intrusion

Section 2.2: UC2 - Building Automation

  • AI-driven HVAC/lighting optimization using occupancy detection
  • BMS integration (Honeywell EBI R410.1) with REST API specifications
  • 3 complete workflows:
  • WF-009: Conference room occupancy → HVAC adjustment
  • WF-010: Cafeteria crowd density → ventilation control
  • WF-011: After-hours occupancy → security alert + local HVAC
  • Infrastructure reuse: Zero incremental hardware (uses UC1 platform)
  • GPU impact: +15-20% utilization (combined total 70-80%)
  • 30-day validation results: 97.7% F1 score occupancy detection

Section 2.3: UC3 - Safety & Compliance Monitoring

  • 3 AI-powered safety functions:
  • PPE Detection: Hard hat + safety vest monitoring (loading dock)
  • Fire/Smoke Detection: Thermal anomaly detection (server room, electrical room)
  • Slip/Fall Detection: Pose estimation for fall incidents (hallways, stairwells)
  • AI Models:
  • YOLO v8n + PPE CNN (28ms inference, 93% accuracy)
  • Thermal processing (CPU-only, 15ms, >85°C threshold)
  • OpenPose pose estimation (45ms inference, 88% fall detection)
  • 2 complete workflows:
  • WF-012: PPE violation → supervisor alert (2-minute response)
  • WF-013: Fire detection → BMS fire alarm + HVAC shutdown (1.5-second response)
  • Hardware: Same UCS XE9305 infrastructure (+10-15% GPU utilization)
  • Integration: BMS fire alarm, ISE badge correlation, ServiceNow, Webex

Section 2.4: Cross-Use Case Correlation

  • Infrastructure sharing model: Single platform serves all three use cases
  • GPU utilization breakdown: UC1 (55-60%) + UC2 (15-20%) + UC3 (10-15%) = 80-95% combined
  • Camera reuse: Same cameras feed multiple AI models simultaneously
  • Model sharing: YOLO v8n person detection reused by UC1, UC2, UC3
  • Network efficiency: 960 Mbps inbound + 58 Mbps outbound = 5% of 20 Gbps capacity
  • Operational benefits: Single management plane, unified monitoring, simplified troubleshooting

Chapter 3: Edge AI Platform Architecture

3.1 Multi-Layer Architecture

  • 5-layer architecture diagram:
  • Layer 1: Camera & Sensor Layer (135 cameras per site)
  • Layer 2: Access Network (Catalyst 9300-48U, PoE+)
  • Layer 3: Distribution Network (Catalyst 9500-40X, routing/aggregation)
  • Layer 4: Edge AI Compute (UCS XE9305 + XE130c M8, NVIDIA L4 GPU)
  • Layer 5: Observability & Integration (ISE, Splunk, TE, AppD, BMS, FTD, XDR, ServiceNow, Webex)
  • Complete latency breakdown: 2ms network + 30ms AI processing + 100ms validation + 100ms actions = 232ms minimum
  • Layer responsibilities and criticality assessment

3.2 AI Model Pipeline

  • Model deployment architecture: ONNX models + Kubernetes (K3s) orchestration
  • Complete inference pipeline (frame-by-frame):
  • Stage 1: Pre-processing (CPU, 5ms) - resize, normalize, colorspace conversion
  • Stage 2: GPU inference (NVIDIA L4, 20ms) - YOLO v8n forward pass
  • Stage 3: Post-processing (CPU, 3ms) - NMS, confidence filtering, zone validation
  • Stage 4: Decision logic (CPU, 2ms) - duplicate check, validation trigger
  • Stage 5: Multi-source validation (parallel, 100ms) - ISE, Splunk, TE, AppD
  • Stage 6: Automated actions (parallel, 80ms) - FTD, XDR, ServiceNow, Webex
  • GPU memory layout: 23.15 GB / 24 GB (96% utilization)
  • CPU core allocation: 18 cores active / 32 cores total (56% utilization)

3.3 Data Flow (Sensor → AI → Action)

  • Complete data flow diagram with timing:
  • T=0ms: Camera frame acquisition
  • T=4ms: Network path (camera → Cat9300 → Cat9500 → edge AI)
  • T=34ms: AI processing complete (pre-process + GPU + post-process + decision)
  • T=134ms: Multi-source validation complete (4 parallel API calls)
  • T=214ms: Automated actions complete (4 parallel actions)
  • T=500ms: Supervisor receives mobile notification
  • Bandwidth analysis:
  • Inbound: 900 Mbps camera streams (4.5% of 20 Gbps capacity)
  • Outbound: 50 Mbps API calls (0.25% of 20 Gbps capacity)
  • Peak burst handling: 1,150 Mbps (5.75% utilization)

3.4 Multi-Site Synchronization

  • Two-site deployment model: Mumbai + Chennai (independent operation)
  • Site independence: Each site operates autonomously during WAN/Internet failures
  • Model synchronization: Blue-green deployment with zero downtime updates
  • Configuration synchronization: GitOps model (FluxCD) with 5-minute reconciliation
  • Update frequency: Quarterly major updates, monthly minor updates, as-needed patches

3.5 High Availability & Failover

  • VRRP failover architecture: Primary/standby nodes per site
  • RTO: <30 seconds (15-sec detection + 5-sec VIP migration + 10-sec container restart)
  • Heartbeat: 5-second interval over 25G midplane fabric
  • Data loss: Minimal (in-flight detections only, event log preserved)

Supporting Documents

Complete Integration Architecture

  • 5-layer architecture diagram with complete data flow
  • 500ms detection timeline with millisecond precision
  • Complete API specifications for all 8 integration points:
  • ISE pxGrid (WebSocket, 50ms)
  • Splunk MLTK (HTTPS REST, 100ms)
  • ThousandEyes (HTTPS REST, 80ms)
  • AppDynamics (HTTPS REST, 90ms)
  • FTD Firewall (HTTPS REST, 50ms)
  • XDR SecureX (HTTPS REST, 40ms)
  • ServiceNow (HTTPS REST, 60ms)
  • Webex Teams (HTTPS REST, 80ms)
  • BMS Honeywell (HTTPS REST, 500ms)
  • Request/response JSON examples for all APIs
  • Bandwidth analysis: 1,010 Mbps total (5% utilization)

Hardware Revision Summary

  • Complete hardware specification for the Cisco Unified Edge platform (UCS XE9305 + UCS XE130c M8)
  • Before/after comparisons for every section

Technical Specifications - Final

Hardware Platform (Cisco Unified Edge)

UCS XE9305 Chassis: - Form factor: 3 RU short-depth (18" / 457mm) - Compute slots: 5 front-accessible (2 used, 3 reserved for Phase 5) - Embedded networking: 2× Unified Edge Management Controllers with 25G Ethernet fabric - Power: Dual redundant PSU (2× 1,000W, N+1), 700W actual consumption (2 nodes) - Cooling: 5× hot-swappable fan modules (N+1), 60dB acoustically optimized - Location: IDF Room, Floor 3 (co-located with Catalyst switches) - Management: Cisco Intersight Infrastructure Service (SaaS, cloud-based)

UCS XE130c M8 Compute Node (Per Node): - CPU: Intel Xeon 6 SoC (32 cores, 2.6 GHz base, 4.0 GHz turbo, 185W TDP) - RAM: 128GB DDR5-4800 (4× 32GB DIMMs, ECC) - GPU: NVIDIA L4 24GB GDDR6 (PCIe Gen5 slot, 72W TDP, 120 TOPS INT8 inference) - Storage: Dual M.2 SATA 512GB RAID 1 (boot) + 2× E3.S NVMe 1TB (event buffer) - Networking: 2× 25G midplane (chassis fabric) + 2× 10G SFP+ front-panel (LACP to Cat9500) - Power: 350W per node average - OS: Ubuntu 24.04 LTS, Docker 26.0, K3s, ONNX Runtime with TensorRT

Deployment: - Mumbai: 2× UCS XE130c M8 nodes (Slots 1-2, primary + standby) - Chennai: 2× UCS XE130c M8 nodes (Slots 1-2, primary + standby) - Total: 4 compute nodes across 2 sites

Camera Infrastructure

Per Site (135 Cameras Total): - 65× Axis P3715-PLVE (Indoor Fixed, 360° panoramic, 6 Mbps avg, PoE+ 25W) - 40× Axis Q6215-LE (Outdoor PTZ, 1080p 32× zoom, 8 Mbps avg, PoE+ 30W) - 20× Axis P1455-LE (4K LPR, license plate recognition, 10 Mbps avg, PoE+ 30W) - 10× FLIR A310f (Thermal, 320×240 thermal, 2 Mbps avg, PoE 15W)

Total Deployment (Both Sites): - 270 cameras (135 Mumbai + 135 Chennai) - 1,920 Mbps aggregate bandwidth (960 Mbps per site) - 6,000W PoE consumption (3,000W per site)

Network Infrastructure

Access Layer: - 6× Catalyst 9300-48U per site (12 switches total) - 48× 1G PoE+ ports, 1,100W PoE budget per switch - 4× 10G SFP+ uplinks (LAG to distribution) - 20 cameras per switch (160 Mbps, 505W PoE = 46% utilization)

Distribution Layer: - 1× Catalyst 9500-40X per site (2 switches total) - 40× 10G SFP+ ports (400 Gbps total capacity per switch) - 2× 10G LAG per edge AI node (20 Gbps per node) - 6× 40G LAG to access switches (240 Gbps aggregate) - 2× 10G WAN uplinks (observability APIs) - Utilization: 0.25% (1,018 Mbps / 400 Gbps)

GPU/CPU Utilization

NVIDIA L4 GPU (24GB per node): - UC1 (Security): 55-60% GPU, 18 GB memory - UC2 (Building): 15-20% GPU, 850 MB memory (reuses UC1 YOLO model) - UC3 (Safety): 10-15% GPU, 4.3 GB memory - Combined: 80-95% GPU utilization, 23.15 GB / 24 GB memory (96%) - Thermal: 65-72°C (within 0-90°C NVIDIA L4 spec) - Power: 70-72W (within 72W TDP envelope)

Intel Xeon 6 SoC CPU (32 cores per node): - Pre-processing: 8 cores (H.265 decode, resize, normalize) - Post-processing: 6 cores (NMS, zone validation, decision logic) - Thermal processing: 2 cores (UC3 fire/smoke detection, CPU-only) - API clients: 4 cores (ISE, Splunk, TE, AppD, BMS, FTD API calls) - Kubernetes + system: 12 cores (K3s control plane, container runtime, OS) - Combined: 18 cores active / 32 cores total (56% utilization) - Peak: 85% during business hours (all use cases active)

Network Latency Breakdown

Camera to Edge AI (4ms total): - Camera → Catalyst 9300 (access): 1ms (1G Ethernet) - Catalyst 9300 → Catalyst 9500 (distribution): 1ms (10G fiber LAG) - Catalyst 9500 → UCS XE130c M8 (edge AI): 2ms (10G fiber LAG)

AI Processing (30ms total): - Pre-processing (CPU): 5ms - GPU inference (NVIDIA L4): 20ms - Post-processing (CPU): 3ms - Decision logic (CPU): 2ms

Multi-Source Validation (100ms parallel): - ISE pxGrid (Mumbai DC): 50ms - Splunk MLTK (NJ DC): 100ms (longest pole) - ThousandEyes (SaaS): 80ms - AppDynamics (SaaS): 90ms

Automated Actions (80ms parallel): - FTD Firewall: 50ms - XDR SecureX: 40ms - ServiceNow: 60ms - Webex Teams: 80ms (longest pole)

Total End-to-End: 4ms + 30ms + 100ms + 80ms = 214ms minimum, 500ms typical (including BMS/FTD actuation + Webex push delivery)