Inspiration

During major natural disasters like earthquakes, cyclones, and floods, centralized cellular infrastructure collapses. Cell towers lose grid power, physical lines sever, and emergency numbers become unreachable. While search-and-rescue teams deploy decentralized sub-GHz mesh radios (e.g., LoRa), these radios face an extreme physical bottleneck: bandwidth. They operate at bytes per second, making voice calls and heavy text paragraphs physically impossible to transmit without crashing the network.

What it does

RelayZero is an offline, edge-AI semantic compression protocol engineered for disaster mesh networks.

Instead of streaming voice or sending lengthy text descriptions across fragile frequencies, RelayZero performs semantic compression at the edge. A responder speaks or types an emergency report into a field unit. A local AI model extracts the vital triage facts on-device and condenses the dispatch into an ultra-dense, 25-byte telemetry packet:
PRIO | LOC | HAZ | CAS | REQ

This delivers an ~88% bandwidth reduction, enabling mission-critical emergency dispatches to easily traverse low-power radio waves where voice calls cannot survive.

How we built it

• Field Unit (Edge Transceiver): Powered by Gemma 4 : E2B running 100% offline via local Ollama. It ingests unstructured voice/text reports and serializes them into structured telemetry tokens with zero internet connectivity.
• Base Station Tactical Command: Built with Python and Streamlit, backed by an asynchronous Flask listener daemon. It deserializes incoming packets, plots dynamic danger perimeters (150m) and incident markers on an offline OpenStreetMap interface, updates live casualty/triage KPIs, triggers browser-synthesized emergency sirens via the Web Audio API, and exports CSV triage logs.
• Offline Mesh Transport: Packets are routed over an offline, zero-internet private socket network structured identically to physical LoRa/Meshtastic packet payloads.
• Development & Prototyping: Scaffolded and optimized using Antigravity for backend architecture and tactical UI styling.

Challenges we ran into

Achieving low-latency local inference on consumer edge hardware without cloud offloading, while ensuring the model strictly and reliably outputs structured, deterministic telemetry tokens from chaotic, natural-language emergency reports.

Accomplishments that we’re proud of

• Achieving true 100% offline end-to-end operation across two separate nodes with zero internet access.
• Shrinking 200+ byte emergency descriptions down to 25 bytes without losing critical triage variables.
• Building a tactical command dashboard with dynamic geospatial hazard rings and synthesized audio sirens.

What’s next for RelayZero

• Porting the transmission layer to native LoRa hardware (ESP32/Meshtastic transceivers over 865/915 MHz).
• Integrating offline edge Whisper models for automatic multi-lingual speech-to-telemetry conversion in remote field operations.