1st Place • Columbia AI for Good Voice AI WebRTC

Patrona: Making "Text Me When You Get Home" Proactive

Feb 28, 2026 4 min read

"Text me when you get home" is something most of us say without thinking. It is a way of expressing concern, but it is also fundamentally reactive. It only works if everything goes according to plan.

At the Columbia University AI for Good Hackathon, we started with a simple question: what would it look like to make that interaction proactive instead of passive?

Interviewing Columbia student
Interviewing Columbia student
We spoke to students across Columbia's campus to understand real late-night safety behaviors.

Patrona is a voice-first safety app designed to stay on a live call as you walk. The idea is not to create a new behavior, but to strengthen one that already exists. People already call friends when they feel unsafe; Patrona makes that support reliable, always available, and responsive.

While the call is active, Patrona keeps light conversation going and tracks response patterns in real time. If a user stops responding beyond a threshold, it performs a verbal check-in. If there is still no response, it escalates and shares live location with emergency contacts.

We also built a safe-word mechanism that allows discreet escalation through the speech pipeline, plus direct voice-triggered emergency actions so users do not need to unlock a phone under stress.

Patrona UI

patrona

Good evening,
Ketaki.

Walk Me Home

Last walk

Today · 11:24 PM · 18 min

Safe
Home
History
Settings
4:32
GPS 40.8075, -73.9626

Listening...

Your companion is right here

Heading to

548 W 113th St, New York

I'm Home
Cancel walk
Alert Active
12:47

Alert sent.

Your contacts have been notified.

Contacts notified

M

Mom

Parent

Notified
S

Princess Leia

Roommate

Notified
Call 911
I'm Safe
End walk

Engineering the Backend: Voice, Silence, and GPS

We used ElevenLabs for real-time voice generation to keep interactions natural, then built reliability around it with voice activity detection, response tracking, and escalation logic tuned for real walking conditions.

I worked across the stack and owned the live GPS pipeline: location pings to Supabase, reverse geocoding through OpenStreetMap, and emergency SMS via Textbelt with human-readable addresses. A real alert looks like this:

Patrona Alert: Katie may need help. Reason: safe word detected. Location: The Amherst, 502, West 110th Street, Manhattan Community Board 7, Manhattan, New York County, New York, 10025, United States Sent by Patrona safety system.

Silence handling was critical. If the user does not respond for 90 seconds, Patrona checks in verbally. Continued silence triggers automatic escalation with the latest reverse-geocoded location to emergency contacts.

The Result

Patrona was built and deployed in 36 hours on Vercel, then won 1st Prize for Best Project Built with ElevenLabs at Columbia AI for Good, along with $5,000 in credits. We are now focused on robustness: threshold tuning, background noise handling, and predictable escalation behavior.

If this is something you would use, we are opening early access here. You can also watch the prototype and user research in our demo here.