An AI safety net for a night nurse who is watching many patients alone — one that keeps working even when the internet, the laptop, and the phone all fail.
Picture a small rural health center at night. One nurse is looking after ten patients. Every few seconds, a device reads each patient's vital signs — heart rate, blood-oxygen (SpO₂), temperature, and breathing rate — and sends them to an AI to be checked.
The AI reads the numbers, compares them against that patient's medical history, and decides how worried we should be:
- normal — nothing to do
- mild — keep an eye on it
- emergency — the nurse (and the doctor) must be told right now
The hard part is not the AI. The hard part is staying alive when things break. Rural internet drops. Laptops die. Phones run out of battery. NeuraRoute is built so that the patient is still being watched no matter what fails.
NeuraRoute keeps a ranked list of AI "brains." For every reading, it uses the first one on the list that is currently alive, and automatically slides down to the next one when something goes offline:
| Order | Brain | Runs when… |
|---|---|---|
| 1 | Cloud AI (the smartest, knows full patient history) | the internet is working |
| 2 | Laptop AI | internet is down, but the laptop is on |
| 3 | Phone AI | the laptop is down too |
| 4 | Tiny chip AI (a small model on an Arduino board) | everything else is gone |
As each brain drops away, the system still works — just with a smaller, simpler AI. This is called graceful degradation: it gets weaker step by step instead of switching off all at once.
Separately from the ladder, the tiny Arduino chip runs a watchdog that checks every reading on its own. If a value crosses a hard danger line (for example heart rate above 135, or SpO₂ below 85), or the small on-chip AI spots a clear emergency, it sends a real notification straight to the doctor's phone — even if the internet, the laptop, the phone, and the main program have all failed.
Built for the Snapdragon Multiverse Hackathon (Bengaluru, July 2026). You can run the entire system on one laptop with no cloud account and no special AI hardware — the AI parts are replaced by a small built-in "mock" so nothing needs to be downloaded.
Install these once:
| Tool | What it's for | Get it |
|---|---|---|
| Python 3.11+ | runs the main program | python.org — on Windows: winget install Python.Python.3.12 |
| Mosquitto | the "post office" the parts use to talk to each other | Windows: winget install EclipseFoundation.Mosquitto · macOS: brew install mosquitto · Linux: apt install mosquitto |
| Node.js 20+ (only for the phone app) | builds the doctor's phone screen | nodejs.org |
| Expo Go (only for the phone app) | the app you open on your phone to see the live screen | App Store / Google Play |
Run these once, in a terminal, inside the folder where you want the project.
Windows (PowerShell):
git clone https://github.com/Sarath-Narendra/neuraroute.git
cd neuraroute
py -3.12 -m venv .venv # create an isolated Python space
.venv\Scripts\Activate.ps1 # turn it on
pip install -r requirements.txt # install the program's needs
cd mobile; npm install; cd .. # phone-app needs (optional)macOS / Linux:
git clone https://github.com/Sarath-Narendra/neuraroute.git
cd neuraroute
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cd mobile && npm install && cd ..This starts everything at once — the "post office," a stand-in AI (so nothing needs downloading), the main program, and all four AI brains on the ladder.
Windows (PowerShell):
powershell -ExecutionPolicy Bypass -File scripts\dev_up.ps1macOS / Linux:
./scripts/dev_up.shWhen it's up, the main program is listening at http://localhost:8080.
Send a made-up critical reading and watch the system flag it as an emergency and fire the alarm.
Windows (PowerShell):
curl -Method POST http://localhost:8080/request -ContentType application/json `
-Body '{"patient_id":"P-03","vitals":{"hr":176,"spo2":79,"temp_c":37,"resp_rate":32}}'macOS / Linux:
curl -XPOST localhost:8080/request -H 'Content-Type: application/json' \
-d '{"patient_id":"P-03","vitals":{"hr":176,"spo2":79,"temp_c":37,"resp_rate":32}}'"Kill" the top brain on the ladder, then send another reading. NeuraRoute will notice and automatically route the next reading to the brain one step down — live.
Windows (PowerShell):
powershell -File scripts\kill_device.ps1 cloud-01macOS / Linux:
./scripts/kill_device.sh cloud-01Windows (PowerShell):
powershell -File scripts\dev_down.ps1macOS / Linux: press Ctrl-C in the window running the stack.
The phone app shows the ten patients, which AI brain is currently in use, and pops up a real alert on every emergency.
-
Make sure your phone is on the same Wi-Fi / hotspot as the computer.
-
Start it:
cd mobile npx expo start -
Open Expo Go on your phone and scan the code that appears.
The app finds the computer automatically. If it can't, you can type the computer's address into the app by hand.
By default the system uses a built-in stand-in AI so anyone can run it. When you have real models and want to use them, NeuraRoute can point each brain at a real model server instead — this is called venue mode. The rest of the system (the ladder, the failover, the phone app, the alarm) does not change; only where the thinking happens changes.
See scripts/README.md and .env.example for the exact settings. In short: copy
.env.example to .env, fill in your keys, and start the model servers with infer_up.ps1
(Windows) or infer_up.sh (macOS / Linux) before bringing the stack up.
A quick map, so you know where to look:
neuraroute/
engine/ the main program — runs the ladder and decides which brain handles each reading
models/ the AI logic: reads vitals, compares to history, returns a severity
servers/ the real per-device model servers (used only in "venue mode")
runtime/ each AI "brain" on the ladder, plus its settings
mobile/ the doctor's phone app
arduino/ the tiny-chip program: the always-on alarm and status light
contracts/ the agreed message formats every part uses to talk
prompts/ the instructions given to the AI
tools/ stand-in AIs so you can run everything with no downloads
scripts/ one-command start / stop / failover helpers
data/ the sample list of ten patients
docs/ setup notes for the real hardware
Detailed technical architecture — component diagram, message flow, and the message formats — will be added here.
Built by:
- Sankarasetty Jaya Abhiram
- Sarath Narendra Kuppala
- Gowtham Sai Kalyan
- Reddy Eswar Anush
Released under the MIT License. All the tools it is built on are open-source.