SafetyBuddy – Hackathon Story

Inspiration

We started with a simple but serious question: what happens when a student is in distress but can’t reach for help?

Panic apps exist, but they depend on someone pressing a button or dialing a number. In a real emergency, that isn’t always possible. That’s what inspired us to build SafetyBuddy — an AI-powered system that can listen, understand, and act automatically, turning a cry for help into immediate support.

What We Learned

Keyword detection alone is not enough. If every time someone said “help” we raised an SOS, supervisors would drown in false alarms.
By combining speech recognition, stress analysis, and emotion/context detection, we were able to filter out noise and focus only on real emergencies.
LLMs became more than just chatbots. We used them to:
- Summarize incidents for supervisors into clear daily/weekly reports.
- Surface patterns across multiple SOS events.
- Highlight the most urgent cases so supervisors could focus where it mattered.

How We Built It

Audio Pipeline
Standardized every recording with ffmpeg.
Extracted stress features (pitch, RMS, tempo) using Librosa. ### AI Models
Whisper → transcribes speech.
DistilRoBERTa → detects emotions like fear or sadness.
DistilBERT → analyzes context (casual phrase vs genuine distress). ### Risk Scoring We created a weighted scoring system that blends all signals: [ \text{Risk} = 30K + 35E + 25C + S ] Where:
(K) = keyword match
(E) = emotion confidence
(C) = context score
(S) = stress weight ### Alert Flow
Log the event in Supabase (transcript, location, risk level).
Send alerts from the student’s verified email to their emergency contacts.
Escalate the case to the supervisor’s Gmail for oversight.
Follow up with the student via an automated email linking to a secure form. ### LLM Analytics
Summarize raw incident logs into human-readable reports.
Provide trend insights (recurring emotions, hotspots, frequency).

- Ensure supervisors act on verified incidents, not noise.

Challenges

False Positives → Early versions triggered on casual “help” mentions; solved by combining multiple signals.
Email Roles → Had to carefully separate student vs supervisor email flows for alerts, escalations, and follow-ups.
Deployment → Installing and running ffmpeg with AI models on Render required Docker/static binaries.

- Performance → Running three AI models + stress detection slowed responses; fixed with chunking and model optimization.

Takeaway

SafetyBuddy became more than just a hackathon project — it proved that AI can be both empathetic and practical.

It listens when students can’t speak for themselves.
It responds in real time with alerts and escalations.
It helps supervisors with clear, actionable insights instead of overwhelming noise. Our biggest takeaway? Technology can be a true safety net, turning scattered distress signals into real intervention and giving students the reassurance that someone is always listening.