Helios

The devastating January 2025 Palisades and Eaton fires exposed a fatal flaw in modern emergency response: centralized networks become single points of failure. When cellular towers burn and the cloud goes dark, static evacuation maps become death traps. We wanted to build a localized, decentralized survival grid that operates entirely on edge hardware. Drawing on prior experience engineering temporal video classifiers for sports analytics and autonomous multi-agent engines, we realized we could fuse real-time street and body camera feeds with macro-satellite data to create a self-healing routing network that thrives offline.

Helios merges dual-stream computer vision with a spatial multi-agent blackboard architecture.

We process real-world disaster datasets and simulated first-responder RTSP body cam feeds using a quantized MobileNetV2 and YOLOv8 pipeline. To solve the edge-compute bottleneck, we implemented a split-inference architecture. The capture device runs only the initial convolutional layers, mapping 32-bit float weights down to 8-bit integers using the standard quantization scheme:

\begin{equation} q = \text{round}\left(\frac{r}{S}\right) + Z \end{equation}

Where r is the real value, S is the scale factor, and Z is the zero-point. These compressed intermediate feature maps are streamed via FastAPI endpoints to a local neighborhood TensorRT base station to complete the segmentation layers, keeping inference latency under 80 milliseconds per frame without a cloud connection.

Visual detections are converted into geospatial polygons and injected into a PostGIS database. We orchestrated three LangGraph agents over this shared state:

The Cartographer Agent subscribes to the PostGIS stream and dynamically penalizes graph edges intersecting with hazard polygons. The updated edge weight w(e) for any road segment e is calculated as:

\begin{equation} w'(e) = w(e) + \alpha \int_{e} H(x, y) \, ds \end{equation}

where H(x,y) is the hazard density derived from the segmentation model.

Navigator Agent: Continuously runs pathfinding algorithms like A^* or Dijkstra over the updated graph, balancing shortest-path extraction with crowd dispersion to prevent localized stampedes. Consensus Agent: Ingests peer-to-peer mobile client data via Bluetooth/Wi-Fi Direct, overriding AI false negatives caused by visual obstructions like black carbon smoke.

The Command Dashboard We built a high-performance command overlay using Next.js and Three.js. This dual-layer interface simultaneously projects macro satellite intelligence (NASA GEDI biomass density) alongside micro ground-level OpenCV bounding boxes, updating the 3D evacuation vectors in real-time as the LangGraph state evolves.

Synchronizing the continuous spatial output of the FastAPI vision stream with the LangGraph state management proved highly volatile. Initially, rapid micro-changes in fire-line detections caused the Navigator Agent to recalculate routes too frequently, leading to graph oscillation. We solved this by implementing a temporal hysteresis threshold, ensuring route updates only trigger when a hazard polygon crosses a critical expansion delta. Additionally, configuring the split-inference feature map streaming required aggressive memory management to prevent memory leaks on the edge nodes during prolonged runtime.

We gained deep exposure to executing AI inference natively on constrained edge hardware, moving away from comfortable cloud API wrappers. Translating theoretical speculative decoding and 8-bit quantization into a functional, low-latency streaming pipeline completely reframed how we approach distributed system architecture. We also learned how to seamlessly bridge geometric spatial databases (PostGIS) with graph-based LLM orchestration, proving that AI can transition from a passive analytical tool into an active, localized survival mechanism.