ShelterSync

Inspiration

During disasters, field volunteers use Slack to report shelter capacity — but messages arrive in different languages, get buried in channel noise, and coordinators can't react fast enough. People get directed to shelters that are already full, wasting critical time and putting lives at risk. We built ShelterSync to fix this.

What it does

ShelterSync is an AI-powered Slack bot that provides real-time shelter capacity management during disaster relief operations:

Multilingual AI Classification: Monitors field channels and uses Groq's Llama 3.3 70B to classify capacity alerts in any language (tested across 13+ languages including Hindi, Bengali, Tamil, Japanese, Spanish, French, and Arabic)
Context Verification: Cross-checks every alert against the last 15 channel messages using AI to filter false positives and duplicates
Smart Redirects: Automatically finds the 2 nearest available shelters using haversine distance and posts a Block Kit redirect card
Proactive Guard: Activates a 30-minute protection window after a shelter is marked full — any message mentioning that shelter triggers an automatic redirect warning
Coordinator DMs: Sends detailed context to the designated coordinator via direct message
Real-time Dashboard: /capacity status shows all shelters, /capacity reopen restores availability
MCP Server: Exposes shelter tools via FastMCP for external AI agent integration

How we built it

Slack Bolt (Python) for event handling, slash commands, and interactive components
Groq AI (Llama 3.3 70B) for multilingual intent classification and context verification
Airtable as the shelter registry and audit log database
FastMCP for exposing shelter operations as MCP tools
Block Kit for rich, interactive Slack messages
pyngrok for secure tunneling to our local development server

Challenges we ran into

Airtable's Single Select fields require options to be pre-created in the UI before API writes succeed — this caused silent failures that took time to debug
OBS Display Capture showed only audio on Windows — switching to Window Capture fixed it
The guard system activated with short names ("shelter 3") but Airtable returned full names ("Shelter 3 - Lajpat Nagar"), causing deactivation to fail. Fixed with partial name matching.
Context verification correctly rejects isolated messages without supporting context — initially looked like a bug but is actually intended behavior

Accomplishments that we're proud of

True multilingual support — not just Indian languages, but any language in the world
The proactive guard system is a unique innovation — no other shelter bot prevents misdirection after an alert
End-to-end pipeline from message to redirect card takes under 3 seconds
Zero false positives in testing thanks to the dual AI verification layer

What we learned

Disaster relief tools must be language-agnostic — volunteers come from everywhere
AI verification (checking alerts against channel context) is essential to prevent alert fatigue
Proactive systems (guards) are more valuable than reactive ones in crisis situations
Slack's Block Kit is powerful enough to build rich operational dashboards