SentinelMesh
AI-assisted drone response: YOLOv8 detection with two local LLM agents arguing before a dispatch. INSPIRON 5.0 (COEP) — AI core lead.
- AI model
- Deterministic rule
- A person reviews
- Not working yet
- inputRecorded video
- AI modelYOLOv8 detection
- ruleConfidence gate
- AI modelTwo LLM agents debate
- outputDispatch on a simulated fleet
- Hackathon build
- YOLOv8
- Two-agent debate
- Simulated fleet
The problem
Every dispatch decision in an incident-response system waits for a human operator, which doesn't scale when several incidents happen at once.
Approach
Route each detection by confidence: clear detections go straight to dispatch, unclear ones go to two local LLM agents — one argues for dispatch, one against — before a decision is made.
Architecture
YOLOv8 incident detection on recorded video; confident detections are dispatched directly and unclear ones go to the two-agent debate (Qwen2.5 via llama-cpp-python, on a local CUDA GPU). A simulated drone fleet with geofencing streams over WebSockets to a live React and Leaflet map, served by FastAPI.
Results and limits
A working end-to-end demo built for the INSPIRON 5.0 hackathon at COEP, on recorded video and a simulated fleet — not tested with real drones.