PrepMTL

Go offline — emergency AI on a USB drive

Pick a USB size, see which emergency AI level fits, and follow the build guide.

A local AI can be wrong. Use it to supplement first-aid cards, trained help and 911 — never instead of them. Download everything while you still have internet. Sizes are approximate; check each download page.

Local AI knowledge pack

Everything this app knows — skills ladder, hazard playbooks, first aid, emergency numbers, kit lists, Montréal locations and the offline levels — bundled into one Markdown file per language for your local AI (Ollama / Open WebUI / llamafile documents folder). It is rebuilt from the app’s current content every time, so new knowledge added to the app reaches the AI automatically.

7 sections · 50 KB per language · content a0923a01

Syncing…

Files: PrepMTL-knowledge-EN.md and PrepMTL-knowledge-FR.md. Point your local AI’s document search (RAG) at that folder.

Your USB drive size

A bigger drive allows a more capable AI plus more maps and library content.

5 GB used55 GB free

Which level is right for you?

Answer three questions. Your level is the lowest of what your devices, comfort, time and USB allow.

Your answers suggest Level 2 — Seed USB

Your device is the limit. To reach Level 3: Laptop with 16 GB RAM.

My offline kit progress

My level: No level chosen yet
Never refreshed

Emergency AI levels

Each level adds to the ones below it — USB space shown is cumulative.

0

Level 0 — Paper & Phone

BeginnerNo USB needed

No AI yet: printed sheets plus an offline library and maps on the phone you already own. Everyone starts here.

Runs on: Any smartphone and a printer.Setup: ≈ 15 min
What it can do
  • Keep contacts, meeting points and medications readable with no power at all.
  • Look up first-aid and medical articles offline (WikiMed).
  • Navigate Montréal and Québec with offline maps.
What it cannot do
  • ✕ Answer questions in plain language — you search and read yourself.
0 of 5 items ready
1

Level 1 — Pocket AI

BeginnerNo USB needed

A small AI model running inside an offline chat app on your phone.

Runs on: Recent phone with 8 GB RAM or more and 8–16 GB free.Setup: ≈ 30 min
What it can do
  • Give quick how-to answers (boil water, shut off gas, keep food cold).
  • Translate short messages EN ⇄ FR and turn notes into checklists.
What it cannot do
  • ✕ Handle complex situations — small models are often wrong.
  • ✕ Replace first-aid cards, a trained person, 911 or 811.
0 of 2 items ready
2

Level 2 — Seed USB

BeginnerRecommendedFits your drive

A tiny assistant plus a first-aid library on a USB key. Runs on almost any laptop, no install.

Runs on: Any laptop with 8 GB RAM, no GPU needed.Setup: ≈ 1 h
What it can do
  • Answer simple how-to questions (boil water, treat a burn, shut off gas).
  • Summarise and translate short texts EN ⇄ FR.
  • Search offline medical and survival articles.
What it cannot do
  • ✕ Reason through complex situations — it makes mistakes often.
  • ✕ Replace first-aid cards, a trained person or 911.
0 of 5 items ready
3

Level 3 — Field USB

IntermediateFits your drive

A capable 3–4B assistant that handles French well, offline maps and a bigger library.

Runs on: Laptop with 16 GB RAM (or a rugged phone with 8–12 GB).Setup: ≈ 2 h
What it can do
  • Hold a short conversation, explain steps in plain language, translate EN ⇄ FR reliably.
  • Answer questions from the offline Wikipedia and medical library you carry.
  • Help plan routes and meeting points with offline maps.
What it cannot do
  • ✕ Give reliable medical dosing or diagnosis.
  • ✕ Read long documents quickly on weak hardware.
0 of 5 items ready
4

Level 4 — Shelter Laptop

IntermediateNeeds ≥ 256 GB

A 7–8B assistant that searches your own documents, on a dedicated emergency laptop or portable SSD.

Runs on: Laptop or mini-PC with 16–32 GB RAM; a GPU makes it much faster.Setup: Half a day
What it can do
  • Answer from your own documents (household sheet, manuals, medical PDFs) with sources shown.
  • Draft messages, checklists and situation reports; translate and summarise longer texts.
  • Be shared by several people, one at a time.
What it cannot do
  • ✕ Guarantee correct answers — always cross-check with the cards or a professional.
  • ✕ Run long on a laptop battery — keep it charged and plan power.
0 of 5 items ready
5

Level 5 — Home Hub

AdvancedNeeds ≥ 512 GB

A low-power mini-PC at home serves AI, library, maps and notes to every phone in the house over local Wi-Fi.

Runs on: x86 mini-PC or ZimaBoard (16–32 GB RAM) + travel router + power station of 500 Wh or more.Setup: A weekend
What it can do
  • Everything in Level 4 for the whole household at once, from any phone browser.
  • Run as one bundled install with Project NOMAD (Kiwix, Ollama AI with document search, offline maps, notes).
  • Keep going through a multi-hour outage on a power station.
What it cannot do
  • ✕ Move easily — it is a home setup, not a go-bag.
  • ✕ Stay safe on a shared network without a password on the AI chat page.
0 of 4 items ready
6

Level 6 — Field Kit (ATU / NOMAD Deck)

ExpertNeeds ≥ 1 024 GB

A portable off-grid station with redundancy everywhere: battery, server, phone backup, LoRa mesh and SDR.

Runs on: Dedicated mini-server (ZimaBoard 2 for endurance, or a faster mini-PC as in the NOMAD Deck) + rugged phone + power station.Setup: Several weekends
What it can do
  • Everything in Level 5 for a household or street, anywhere, from one case.
  • Keep working when the grid is down: battery, optional solar, mesh text over LoRa.
  • Monitor radio (weather, aircraft, FM/AM) and keep a backup copy of the knowledge on the phone.
What it cannot do
  • ✕ Run large models fast on a ZimaBoard without an added GPU — expect a few words per second.
  • ✕ Transmit on the SDR legally without an amateur radio licence, or guarantee LoRa range.
0 of 5 items ready

Build your USB drive

  1. Choose the drive and format it

    Use a USB 3.x drive or portable SSD. Format as exFAT so Windows, macOS and Linux can all read it and files over 4 GB are allowed (FAT32 cannot). Label it PREPMTL-OFFLINE.

  2. Create the folder layout

    One folder per kind of content keeps it easy to find under stress. Add a README.txt with the three steps to start.

    PREPMTL-OFFLINE/
      README.txt
      01-ai-models/
      02-ai-runtime/
      03-library-kiwix/
      04-maps/
      05-documents/
      06-installers/
  3. Download while you have internet

    Fetch the items listed for your level into the matching folders. Check file sizes against free space first, and keep the download page URLs in README.txt.

  4. Verify the files

    Create checksums so you can detect a damaged copy later.

    cd /path/to/PREPMTL-OFFLINE
    sha256sum 01-ai-models/* 03-library-kiwix/* > checksums.sha256
    sha256sum -c checksums.sha256
  5. Start the AI (no install, no internet)

    Level 2: run the single-file llamafile, then open the address it prints. Windows cannot run a program file over 4 GB, so only small models work as one .exe. Level 3+: keep the program and the model separate — install Ollama from 06-installers and point it at the model folder, or run the small llamafile program with the model file passed to it.

    # Level 2 — macOS / Linux
    chmod +x 01-ai-models/model.llamafile
    ./01-ai-models/model.llamafile
    # Level 2 — Windows (file under 4 GB): rename to model.llamafile.exe and double-click
    
    # Level 3+ — program and model kept separate
    02-ai-runtime/llamafile.exe -m 01-ai-models/model.gguf
    # or with Ollama
    OLLAMA_MODELS=/path/to/PREPMTL-OFFLINE/01-ai-models ollama serve
  6. Open the offline library

    Open any .zim file in the Kiwix app, or serve them to the whole local network.

    kiwix-serve --port 8080 03-library-kiwix/*.zim
    # then browse http://<this-computer>:8080
  7. Test in airplane mode and refresh twice a year

    Unplug the internet and try one question, one map and one article. Replace models and ZIMs every 6 months and re-run the checksums.

Advanced — ATU briefcase hardware stack

A portable, low-power kit: battery, fanless mini-server, rugged phone, LoRa mesh radio, SDR, flashlight and a 3D-printed case. Based on the Auxiliary Tech Unit v1 notes (UrbanCircles).

Hardware parts (tick what you will run)

  • ZimaBoard 2~10 W
    Fanless server: AI, Kiwix, maps, chat.
    Intel N150 quad-core, 8–16 GB RAM, 32–64 GB eMMC, 2× SATA, 2× 2.5 GbE, PCIe 3.0 x4. About 6 W idle. No built-in Wi-Fi.
  • Wi-Fi access point~4 W
    Gives the server a local wireless network.
    Any small travel router or USB Wi-Fi adapter.
  • AGM G3 Pro~4 W
    Field interface and independent backup (maps, local AI, Kiwix, thermal camera).
    6.72" 120 Hz, 12 GB RAM, 512 GB, 10,000 mAh, IP68/IP69K, MIL-STD-810H.
  • Heltec LoRa 32 V4~0.5 W
    Meshtastic node: text messages without cell towers.
    LoRa up to 28 dBm, Wi-Fi, Bluetooth, optional GNSS, 18650 battery, solar port. Range depends on terrain and local rules.
  • PortaPack H4M~1.5 W
    Standalone SDR receiver: weather, aircraft, FM/AM, sub-GHz.
    1 MHz–6 GHz, touchscreen, Mayhem firmware, USB-C charging.
  • Wuben X1 Pro~0 W
    Light, signalling and 15 W reverse charging.
    Up to 12,300 lm, 410 m throw, 2× 21700 cells (9,600 mAh), IP65, lockout mode.

Power calculator

Average draw
19 W
Estimated runtime
4.2 days

To last 7 days you must average ≤ 11 W (or add solar).

Assumes 90% inverter efficiency. A 30 W average on 2,048 Wh is about 2.8 days before losses (≈ 2.5 days after), not a week — measure real watts.

How redundancy works

Server and phone both provide knowledge and AI. Flashlight and power station both provide charging. Phone and LoRa nodes both provide communications. SDR, phone and server each add different situational awareness.

LoRa range (about 10 km) depends heavily on terrain, antenna placement, obstructions and legal power limits in your region (Canada: 915 MHz band).

Build steps

  1. 1. Size the power first

    Add up the average watts of everything you will run, then divide the battery watt-hours by that number (use the calculator). The source design assumes about 30 W, which gives roughly 2.8 days on a 2,048 Wh station — not a week. For a week, aim for about 12 W average or add solar.

  2. 2. Prepare the server

    Install ZimaOS or Ubuntu Server on the ZimaBoard 2 while online. Add an NVMe or SATA SSD (1 TB) for the library. Install Ollama, Open WebUI and kiwix-serve as containers, then save the images so a rebuild works offline.

    docker save ollama/ollama ghcr.io/open-webui/open-webui kiwix/kiwix-serve -o images.tar
    # restore later with: docker load -i images.tar
  3. 3. Give it a local network

    The board has no Wi-Fi: connect a small access point by Ethernet and set a fixed address. Phones and laptops then open the chat and library in a browser with no internet.

  4. 4. Set up the phone as backup

    Copy a Level 3 model, offline maps and Kiwix to the phone’s storage so it still works if the server fails. Pair it with the LoRa node over Bluetooth in the Meshtastic app.

  5. 5. Configure the mesh radios

    Flash Meshtastic, pick the correct region (Canada uses the US 915 MHz band), set a private channel key and test two nodes outdoors. Range varies a lot with terrain and antenna height; the roughly 10 km quoted is a best case.

  6. 6. Print or build the case

    The ATU v1 case is a 3D-printed briefcase (STEP files published by its creator). Use PETG rather than PLA for impact and heat resistance, and adjust tolerances to your printer. Any rugged case with foam works too. Add a printed card with frequencies, Morse code and your household sheet.

  7. 7. Rehearse the outage

    Run the whole kit for 24 hours with the grid unplugged. Measure real watts with a plug-in meter, compare with the calculator, and fix the gaps. Write the start-up steps on the case lid.