PAN!C

PAN!C exists to ensure that no one faces a federal enforcement encounter alone. PAN!C alerts your people, shares your location, and puts a calm, knowledgeable AI assistant in your hands

Comment 1

What it does

PAN!C is an emergency response app built for immigrant communities facing ICE enforcement encounters. With a single button press, PAN!C captures your GPS coordinates, generates a unique incident ID, and simultaneously dispatches SMS alerts to every emergency contact on file. A built-in AI assistant - powered by Google Gemini - explains your legal rights in real time, responds in your language, and remembers your situation across sessions. Users can also securely store critical documents (IDs, visas, legal paperwork) with AES-256 encryption, set up automated check-ins so contacts are alerted if you go silent, and disarm the panic alert with a personal safe phrase once they are safe.

How we built it

PAN!C is a full-stack platform built across web and mobile. The frontend is React + TypeScript + Vite, with a React Native + Expo mobile app. The backend is Node.js + Express, backed by MongoDB Atlas for persistent storage and JWT-based authentication.

The AI layer combines Google Gemini 2.5 Flash with Backboard.io for long-term semantic memory - before every Gemini call we search past sessions and inject relevant context into the system prompt, so the assistant remembers prior incidents, language preferences, and conversation history across sessions. Emergency notifications are delivered via Twilio SMS and Firebase Cloud Messaging. Documents are encrypted with AES-256-GCM before being stored in the database.

Challenges we ran into

Getting AI memory to work without slowing down emergency responses was our hardest technical challenge. Gemini alone has no cross-session memory, so we built a semantic retrieval layer with Backboard - searching relevant memories before every API call and saving exchanges asynchronously so they never block the response.

We also hit Google API quota limits mid-build, forcing a rapid migration to a paid tier and a model version update under pressure. Getting Expo Go to reliably connect across different network configurations - particularly with Tailscale VPN affecting local routing on shared hackathon WiFi - required deep debugging of how Metro bundler binds to network interfaces.

Accomplishments that we're proud of

We are proud of building a genuinely functional emergency tool — not a prototype, but a working app with real SMS delivery, real encrypted document storage, real AI responses, and persistent memory that survives across sessions. We are proud that the AI assistant actually knows who you are the second time you open it. And we are proud of building something that could matter in a real person's worst moment.

What we learned

We learned that emergency tools have to work perfectly or not at all

  • there is no "good enough" when someone's safety is on the line. We learned how to architect a parallel alert system that fires multiple services simultaneously under a single trigger. We learned how to give an LLM genuine long-term memory using semantic search and prompt injection. Most importantly, we learned that building for vulnerable communities means earning trust through reliability, privacy, and simplicity in every single decision.

What's next for PAN!C

We want to expand the AI assistant's legal knowledge with jurisdiction-specific rights information across all 50 states. We plan to add a community network feature so users can silently alert nearby PAN!C users during an enforcement sweep. . Long term, PAN!C should be a platform that community organizations can deploy and manage for the people they serve.

Share this project:

Updates