RelayZero: Offline Edge-AI Semantic Compression for Disaster Mesh Networks

CS KARTHIK — Thu, 08 Oct 2026 10:58:02 +0000

The Problem: When the Grid Goes Dark

When catastrophic natural disasters strike, the first thing to collapse is centralized telecommunications. Cell towers lose power, fiber lines are severed, and dialing emergency numbers fails because there is no base station to route the call.

In these blackouts, search-and-rescue teams rely on decentralized, sub-GHz mesh radios (like LoRa). These radios are incredible: they run for days on tiny batteries, require no central towers, and their low frequencies punch straight through concrete rubble.

But there is a fatal physical bottleneck: bandwidth.

Mesh radios transmit at bytes per second, not megabits. If you try to force a standard voice call (thousands of bytes per second) or even a lengthy text paragraph through a mesh network, the frequency chokes, packets drop, and the entire network crashes.

What We Built: RelayZero

RelayZero is a 100% offline, edge-AI semantic compression protocol. We built it to bridge the gap between human panic and low-bandwidth machine telemetry.

Instead of sending messy, paragraph-long emergency dispatches over fragile radio links, RelayZero intercepts the message at the edge. It uses Gemma 4 : E2B to strip away conversational filler, extract the vital triage facts, and compress the dispatch into a dense 25-byte telemetry packet formatted as:
PRIO:CRITICAL | LOC:LIBRARY | HAZ:FIRE | CAS:2 | REQ:MEDICS

This intelligent edge compression yields an ~88% reduction in data size, allowing life-saving SOS alerts to slip effortlessly through low-power mesh radios.

How We Built It

We designed a two-node, 100% offline architecture to simulate a disaster deployment zone. To quickly scaffold our Python backends and style our tactical UI interfaces, we leveraged Antigravity.

The Field Unit (Laptop A - Edge Transceiver)
The AI Engine: We ran Gemma 4 : E2B locally via Ollama. It requires zero internet access.
The Workflow: A frantic responder speaks or types into the Python UI (transmitter.py). Gemma parses the raw input locally, extracting the core variables (Priority, Location, Hazard, Casualties, Requirements) and serializing them into the 25-byte string.
The Base Station (Laptop B - Tactical Command)
The Dashboard: Built using Python and Streamlit (base_station.py), featuring a daemon Flask background listener on port 5000 to catch incoming packets.
Geospatial Triage: Incoming coordinates are instantly mapped onto an offline OpenStreetMap via Folium, drawing a dynamic 150-meter danger perimeter around the incident.
Zero-Network Alerts: For packets flagged as CRITICAL, the frontend automatically synthesizes a two-second police siren using the browser's native Web Audio API (meaning no external audio file downloads are needed).
Post-Disaster Logging: Triage officers can export the real-time queue to a structured CSV for post-operation forensics.
The Transport Layer
Due to hackathon constraints, we simulated the physical mesh layer over a private, zero-internet local TCP socket network (172.20.10.2:5000). However, the 25-byte string payload is byte-for-byte identical to what would be passed into a physical LoRa transceiver over UART.

See it in Action

GitHub Repository: https://github.com/CSKarthik-0/hacktoberfest-hack-day-coimbatore-x-init-club-and-idea-club
Live Video Demo: https://youtu.be/5DZY78ddo0E

DEV Community: CS KARTHIK

RelayZero: Offline Edge-AI Semantic Compression for Disaster Mesh Networks