EdgeMind: 5G-MEC Intelligence Orchestration
![]()
Real-time AI orchestration at telecom edge with Strands agent swarms. Brings intelligence to 5G Multi-access Edge Computing sites for sub-100ms decision making.
Project Overview
Built a 5G-MEC (Multi-access Edge Computing) orchestration system that deploys Strands-based multi-agent swarms directly at telecom edge sites. The system monitors local metrics, detects performance degradation, and self-orchestrates routing and resource decisions without cloud dependence.
Today's AI systems trade speed for intelligence. Edge devices process fast but lack complexity; the cloud processes deeply but adds latency. For real-time applications—autonomous vehicles, industrial control, or competitive gaming—milliseconds matter.
EdgeMind brings intelligence to the edge through threshold-based orchestration, MEC-native intelligence, and swarm coordination—all achieving sub-100ms routing decisions.
Key Innovation
Threshold-Based Orchestration
- Monitors latency, CPU/GPU load, and queue depth
- Triggers intelligent swarm responses automatically
- Adapts to network conditions in real-time
- No cloud dependency for critical decisions
MEC-Native Intelligence
- Strands agents deployed directly at telecom edge sites
- Located near RAN (Radio Access Network) controllers
- Complete agent set at each MEC site
- Local MCP tools for metrics and operations
Swarm Coordination
- Agents collaborate across MEC sites
- Balance load without cloud involvement
- Autonomous failover between sites
- Consensus-based decision making
Real-Time Performance
- Sub-100ms routing decisions
- 99.9% availability through redundancy
- Autonomous load balancing
- Intelligent adaptation to conditions
Architecture & Technical Approach
System Flow
User Devices (5G)
→ MEC Site A (Primary)
→ Swarm Coordination
→ MEC Sites B & C (Fallback)
→ AWS Cloud (Passive Observer)
MEC Site Components
Each MEC site contains:
Complete Strands Agent Set:
- Orchestrator Agent
- Load Balancer Agent
- Resource Monitor Agent
- Decision Coordinator Agent
- Cache Manager Agent
Local MCP Tools:
- metrics_monitor
- container_ops
- inference_engine
- telemetry_logger
- memory_sync
Agent Architecture
| Agent | Role | Deployment |
|---|---|---|
| Orchestrator Agent | Threshold monitoring & swarm triggering | MEC Site Controller |
| Load Balancer Agent | Distribute workload across MEC sites | Strands Swarm Member |
| Resource Monitor Agent | Track CPU/GPU/latency metrics | Strands Swarm Member |
| Decision Coordinator Agent | Coordinate swarm consensus | Strands Swarm Member |
| Cache Manager Agent | Local model and data caching | Strands Swarm Member |
Business Use Cases
Gaming & Esports
- Real-time NPC dialogue: Device SLM for instant responses
- Game state analysis: MEC swarm coordination for regional multiplayer
- Performance analytics: Cloud observability (passive)
- High GPU usage: 85-95% utilization for rendering and AI
Autonomous Vehicles
- Collision detection: Device SLM for ultra-low latency safety (<30ms)
- Traffic coordination: MEC orchestrator manages regional traffic flow
- Fleet analytics: Cloud monitoring and long-term insights
- V2X communication: Vehicle-to-everything coordination
Smart Cities & IoT
- Sensor processing: Device SLM for immediate responses
- City-wide coordination: MEC swarm balances infrastructure load
- Urban planning: Cloud analytics from aggregated MEC data
Healthcare
- Patient monitoring: 50-200 patients per MEC site
- HIPAA compliance: Local processing for privacy
- Medical alerts: Real-time critical event detection
Technology Stack
- Edge Agents: Strands framework with Claude 3.5 Sonnet integration
- AI Model: Claude API for real agent coordination
- MEC Infrastructure: Docker/Kubernetes on edge compute nodes
- Dashboard: Streamlit with real-time simulation and dual-mode operation
- Orchestration: Threshold-based swarm coordination with MCP tools
- AWS Integration: AgentCore Memory + Orchestration only (passive)
- Communication: Direct MEC-to-MEC networking
Live Dashboard Features
Dual-Mode Operation
- Mock Data Mode: No API key required, simulated agents
- Real Strands Agents Mode: Full Claude API integration
Real-Time Metrics
- Latency (ms) — target <100ms
- CPU Usage — trigger >80%
- GPU Usage — monitoring utilization
- Queue Depth — request backlog
Swarm Visualization
- Green: Healthy MEC sites
- Red: Overloaded sites
- Gray: Failed sites
- Lines: MEC interconnections
Agent Activity Stream
- Info: Normal operations
- Success: Consensus achieved
- Warning: Threshold breach
- Error: System failure
Enhanced Demo Scenarios
Gaming: High GPU usage (85-95%), multiplayer synchronization, NPC AI processing
Automotive: Ultra-low latency (<30ms), safety-critical systems, V2X communication
Healthcare: Patient monitoring (50-200 patients), HIPAA compliance, medical alerts
Normal: Balanced resource utilization and standard MEC operations
Automated Demo Mode
- Cycles through all scenarios every 15 seconds
- Scenario-specific metrics and thresholds
- Enhanced visualizations with context-aware indicators
- Start/Stop controls for presentation mode
Skills Demonstrated
5G/MEC Architecture: Multi-access edge computing, RAN integration, telecom infrastructure, edge deployment
Multi-Agent Systems: Strands framework, swarm coordination, consensus algorithms, agent orchestration
Real-Time Systems: Sub-100ms latency targets, threshold-based triggering, performance optimization
Distributed Systems: MEC-to-MEC networking, failover mechanisms, load balancing, redundancy
Cloud Integration: AWS AgentCore, passive observability, hybrid edge-cloud architecture
Dashboard Development: Streamlit, real-time visualization, dual-mode operation, automated demos
DevOps: Docker/Kubernetes, edge deployment, monitoring, telemetry
MCP Protocol: Tool design, metrics monitoring, container operations, inference engines
Expected Outcomes
- Sub-100ms decision making for real-time applications
- Autonomous load balancing without cloud dependency
- 99.9% availability through MEC site redundancy
- Intelligent swarm coordination adapting to network conditions
Future Work: ICEO Framework
The next phase extends toward ICEO (Intelligence-Centric Edge Orchestration), where each MEC site acts as a learning agent within a distributed intelligence fabric.
Planned Research:
- Multi-MEC simulation for latency and consensus testing
- Reinforcement-based learning between edge and cloud layers
- Formalize and publish ICEO as a framework for autonomous 5G orchestration
Links
- GitHub: mec-inference-routing
- Live Demo: Interactive Dashboard
- Architecture Guide: Full Documentation
- Demo Scenarios: Enhanced Scenarios