Inspiration

When a disaster hits, help lines don't fail because nobody can rank requests. 911 and field triage already do that. They fail because of silence. Someone texts "Mom is 82, on oxygen, the power's out," the message sits in a queue, and by the time anyone reads it, that person has stopped answering. An unanswered message looks exactly like a case that's been handled.

State Farm's track, The Worst Day, at Scale, made us ask one question about every feature: does it help make sure nobody is overlooked? That became the product. Everyone ranks requests. We chase the people who stop answering.

What it does

  • We text every household in the storm's path. Nobody has to be the one who calls for help. Reply 1 and you're marked safe. Reply with your situation and it becomes a ranked report.
  • Silence escalates. No reply means a re-text, then a phone call, then a red urgent flag for a human to go look. Claiming a case stops the check-ins, so nobody gets chased after help is on the way.
  • The call is automated too. A voice asks if they're safe. Press 1 and they're marked safe. Press 2 and they choose: describe what happened, which files a ranked report in their own words, or have someone sent. A hang-up or silence leaves them urgent, so nobody is dropped.
  • Anyone can also text in directly. Each message is turned into facts in the person's own words (age, medical needs, hazards, how many people) and ranked by readable rules. Every reason shows the words that triggered it.
  • Responders share one live board. Only one person can claim a case, so two people never drive to the same door.

How we built it

Three pieces, nothing else:

  1. A Spacetime module (TypeScript) is our database and our backend: requests, households, storms, check-ins, an outbox work queue and call attempts. The ranking rules and the escalation ladder are reducers running inside the database, not a server beside it. A scheduled tick fires every second and moves cases through the state machine. Claiming is a single transactional reducer, so in a two-tab race exactly one responder wins.
  2. A React board on Vercel subscribes to the tables with useTable, so ranks, counts and red flags update live with no polling and no API layer.
  3. A Node bridge sends and receives iMessages through Photon, places voice calls through ElevenLabs, calls Claude to pull structured facts and quotes out of each message, and works through the outbox: extract_facts, send_text, place_call.

Challenges we ran into

  • Photon can't text first. On the free plan, the agent can only message someone who has already messaged it, and US SMS needs carrier approval that takes days. So real phones opt in once by texting JOIN (which is how storm alerts work anyway), and the simulated households answer through seeded randomness on the server.
  • There's no "place a call" API. Voice goes through SIP: ElevenLabs places the call over Photon's trunk. Getting the agent to stay on script took several real calls. It repeated its own lines and kept talking after the report was filed. Free call minutes are scarce, so every test counted.
  • One shared live database. Before we locked it down, someone hit an unprotected button and injected 20,000 rows. Destructive and demo reducers are now admin-only on the server.
  • Two bridges mean double messages. If two teammates ran the bridge at once, both polled Photon and every message arrived twice. Lesson: exactly one bridge, owned by whoever watches its log.
  • Four people plus their AI assistants in one repo. Agents happily rewrote whole boards, and two redesigns collided. We switched from tickets to owning areas and merged small PRs often.

What we learned

  • Putting logic in the database (Spacetime reducers plus subscriptions) removed a whole server layer and made the claim race a non-problem.
  • Explainable rules beat a smarter model for trust: a responder can see why a case is ranked where it is. Claude only extracts the facts; our rules set the level.
  • Real-world constraints (opt-in messaging, carrier rules, call costs) shaped the design as much as the code did.
  • Guardrails come first on a shared database, not after the first incident.

What's next

A registered SMS number so anyone can text in without iMessage, handing urgent cases to dispatch directly, and load testing the board with thousands of storm households at once.

Built With

Share this project:

Updates

Submission history