Chapter 3: Edge AI Platform Architecture¶
Multi-Layer Architecture & AI Model Pipeline
This chapter details the complete platform architecture for Abhavtech's Edge AI deployment, including the five-layer architecture design, AI model pipeline, data flow from sensor to action, and multi-site synchronization.
Chapter Content¶
The platform architecture is delivered as a comprehensive single document covering:
Complete Platform Architecture¶
Section 3.1: Multi-Layer Architecture
Five-layer design: Hardware Layer, Container Orchestration Layer, AI Processing Layer, Integration Layer, and Application Layer.
Section 3.2: AI Model Pipeline
Model training, ONNX export, TensorRT optimization, container registry (Harbor), and automated deployment.
Section 3.3: Data Flow (Sensor → AI → Action)
Complete data flow from camera/sensor input through GPU inference, multi-source validation, decision logic, and automated action execution.
Section 3.4: Multi-Site Synchronization
Model synchronization between Mumbai and Chennai, configuration management, and centralized model registry.
Section 3.5: High Availability & Failover
VRRP configuration, 2-node active-standby setup, <30 second failover RTO, and disaster recovery procedures.
Platform Highlights¶
- Compute: Cisco UCS XE9305 chassis + UCS XE130c M8 compute nodes (Intel Xeon 6 SoC, 32 cores, 128GB RAM)
- GPU: NVIDIA L4 24GB (120 TOPS INT8, 20ms inference time, 72W TDP)
- Orchestration: Kubernetes 1.28+ with GPU device plugin
- Container Registry: Harbor (private registry for AI models)
- Storage: 2TB NVMe (7-day event buffer, 250GB/day retention)
- Network: 2× 10G LAG uplinks per node (20 Gbps, 5% utilization)
Architecture Philosophy¶
Distributed Intelligence + Centralized Wisdom:
- Edge: Real-time AI inference (<500ms latency), 7-day local buffer, quick decisions
- Centralized: Model training, long-term storage (Splunk 90-day retention), complex correlation
Previous: Chapter 2: Use Case Architecture | Next: Chapter 4: Integration Architecture