TheWeb

One SMS number + OpenAI local model = A local emergency broadcast system. By texting the number, the LLM can identify the emergency and broadcast it to every subscriber in the postcode to alert.

Comment

Inspiration

Emergencies outrun official alerts. When there’s a fire, a gas leak, a crash, or someone missing, neighbors need a way to tell each other now—without apps, logins, or Wi-Fi. The Web is a “neighborhood pager”: one phone number + SMS + a small local LLM that turns raw texts into instant, hyper-local alerts.

What it does

  • Anyone joins by texting a short oath; then they can report incidents by SMS.
  • A local language model classifies each message (spam vs. real), asks only for missing pieces (what/where/when), and when complete, composes a concise public alert.
  • The alert is broadcast by postcode to nearby subscribers. Every alert includes a safety note and stays under SMS limits.
  • Works fully offline in demo mode; can use Twilio (A2P 10DLC needed for unrestricted sending) or a USB GSM modem for real SMS.

How I built it

  • Stack: .NET 8 console app + SQLite.
  • Onboarding: strict SMS “oath” parser so join is human-simple yet machine-reliable.
  • LLM brain: OpenAI-compatible local endpoint (e.g., LM Studio). I send only the last few messages from the same reporter (within 30 minutes), extract hints (postcode/time/location), and ask the model for one minified JSON decision.
  • Broadcast: when complete, I format a short alert and SMS it to subscribers in the target postcode, excluding the reporter.
  • Safety/ops: 30-minute cooldown per reporter, inbound de-dupe, and a demo replay script for predictable demos.

Challenges I ran into

  • SMS realities: needing A2P 10DLC registration (otherwise “verified numbers only”) made true end-to-end testing tricky.
  • LLM reliability: getting strictly valid JSON under token limits; preventing the model from giving unsafe advice; keeping follow-ups ultra-short.
  • Partial info: robustly extracting postcode/time/location from messy texts and threads, not just the last message.
  • Race/dup: avoiding double processing when polling inboxes; enforcing a per-reporter cooldown so alerts don’t spam.

Accomplishments that I’m proud of

  • A working, end-to-end SMS → LLM → geofenced broadcast pipeline that runs locally.
  • A clean demo mode so anyone can try it without a provider or hardware.
  • Human-friendly UX in pure SMS: a one-line oath to join, one-line follow-ups for missing info, and clear, short public alerts.

What I learned

  • For crisis tools, SMS first wins: zero friction beats any fancy app.
  • LLMs behave best with strict schemas + guardrails (seed JSON, extract only the first balanced object, post-validate).
  • Design for failure paths (timeouts, 400/429s, A2P constraints) from day one—your demo will thank you.

What’s next for The Web

  • Go fully offline: USB GSM/LTE modem + SIM (no cloud dependency).
  • Stronger signal: incident clustering + simple “2-reporter” corroboration gates; lightweight YAML playbooks for severity.
  • Ops panel: minimal web UI to review/override alerts, see heatmaps, and adjust templates.
  • A2P 10DLC & onboarding: register properly, add multilingual flows, and polish the safety copy.
  • Open source packaging: one-command setup, sample replay scripts, and a tiny local model preset so anybody can run a neighborhood pager.

Built With

Share this project:

Updates