Inspiration
When hurricanes, floods or wildfires hit, power and cell networks are often the first things to fail, right when people need help most. Responders and volunteers rush in, but without a shared picture of who needs what, coordination quickly breaks down and people get missed. We wanted to make sure everyone can still call for help and be found, even when the infrastructure is gone.
What it does
RescueMesh is a low-cost, fully offline emergency network.
- Send an SOS without signal: A person plugs our ESP32 + LoRa radio into their phone and uses our app to send an emergency report: severity, situation, location and phone battery level.
- Local triage: The report reaches a town hub, where a locally running Nemotron agent categorizes and prioritizes it for a human dispatcher.
- Follow-ups: The agent can message the person on its own to gather more details, so no one is left in the dark.
- Regional coordination: Town hubs report their status to a central hub. Agents talk to each other through very short compressed messages to spot overloaded hubs and recommend reallocating dispatchers from neighboring towns.
- Humans stay in control: Agents gather information and make recommendations, but only dispatchers can reorder or close emergencies.
How we built it
- Hardware: ESP32 boards with LoRa radios and extended antennas for long-range, low-power communication.
- Phone app: Swift app that builds the emergency report and sends it through the ESP32.
- Hubs: Each hub runs its own backend in MongoDB with a lightweight Nemotron model running locally, so nothing depends on the internet.
- Compression: Emergency data is encoded into binary fields and compressed to minimize LoRa airtime, and agents communicate with very short text.
- Dashboards: A dispatcher view for each town hub and a regional view for the central hub.
Challenges we ran into
- Fitting useful emergency data and agent-to-agent communication into LoRa's very limited bandwidth.
- Running AI agents fully offline on lightweight hardware instead of relying on cloud models.
- Designing agents that are useful on their own while keeping critical decisions with human dispatchers.
- We noticed a main issue with sending signals back from the main hub to personal transponders and we worked through increasing return signals if the transaction was not complete yet
Accomplishments that we're proud of
- A working end-to-end path from a phone with no signal, over LoRa, to a hub where Nemotron triages the emergency.
- Agents that coordinate across hubs using tiny compressed messages instead of an internet connection.
- In testing, we achieved up to half a mile with an expected range of 2 miles with our prototype. The expected range given full hardware and software is around 5-10 km
What we learned
- How to work within the constraints of low-bandwidth radio and design data formats around them.
- How to build agentic workflows that go beyond a chatbot: agents that plan, act and coordinate, with clear limits on what they can decide.
What's next for RescueMesh
- Replace the simulated hub-to-hub network with real LoRa and satellite links.
- Add message authentication to block spoofed or fake SOS reports.
- Solar and battery-powered hubs for long-term deployments.
- Field testing with local emergency responders and community emergency response teams.
- Pre-staged RescueMesh kits at fire stations, shelters and community centers.
