SentinelMesh
Autonomous drone fleet that detects, debates, and dispatches — without human input.
The Problem
Urban safety incidents require response in seconds, not minutes. Existing systems rely on human operators for every dispatch decision — a bottleneck that scales poorly under simultaneous multi-zone incidents.
Approach
Adapted the MAPE-K autonomic computing loop from server healing into drone fleet management. Added an Adversarial Debate Engine between detection and dispatch to eliminate false positive responses.
Architecture
Split inference across two machines: Garuda Linux (RTX 3050 Ti) handles all CUDA inference — YOLOv8n at imgsz=480 and Qwen2.5-1.5B-Instruct Q4_K_M via llama-cpp-python. Windows machine handles FastAPI, fleet simulation, and React + Leaflet.js dashboard. Confidence tier gating bypasses LLM debate for high-confidence detections (≥ 0.88), cutting dispatch to ~50ms. Ghost Protocol enables automatic drone reassignment on failure with zero operator intervention.
Results
55fps real-time detection on consumer edge hardware. Sub-50ms dispatch for high-confidence events. Self-healing fleet with zero-downtime reassignment. Predictive pre-positioning via Temporal Shadow Deployment.
