SafeWalk

Inspiration

The idea for SafeWalk came from a feeling most of us have had — that quiet unease of walking home alone at night, hoping nothing goes wrong. One of our team members recalled a late-night walk back from campus where they kept their phone open just to feel like someone was there. That moment stuck.

Existing safety tools fall into two camps: passive location sharing or panic buttons you press when it's already too late. Neither solves the real problem — the long, vulnerable window before anything goes wrong. We wanted to build something that stays with you. Not a tool you reach for in a crisis. A presence that walks alongside you.

What it does

SafeWalk is an AI-powered walking and driving companion app with a live safety monitoring layer. At its core, the experience is a two-way AI conversation — like being on a phone call with a calm, attentive companion the entire time you're moving.

** AI Voice Companion ** - a real two-way conversation powered by ElevenLabs, keeping you engaged and present throughout your entire trip
Route Awareness — 2–3 Google Maps route options with short AI-generated contextual observations per route
Live Safety Status — a visible indicator reflecting real journey state
Check-In Prompts — if the user goes quiet or deviates from route, the app asks: "Hey, are you okay?"
Danger Word Trigger — a configurable word that immediately escalates
Manual SOS — always visible, always one tap away
Trusted Contact Alerts — Twilio SMS with live location link, trip context, and alert reason

How we built it

Layer	Technology
Mobile Frontend	React Native + Expo (TypeScript)
Backend Runtime	Node.js + Express.js (TypeScript)
Data Storage	In-memory `Map` — no database, MVP-ready
Maps & Routing	Google Maps Platform (Directions, Geocoding, Places)
AI Voice	ElevenLabs
Alerts	SMTP

We split the team across four tracks — frontend, backend logic, AI/voice integration, and alerts/demo — and worked in parallel, syncing on a shared API contract so both sides could build independently before connecting.

Challenges we ran into

Scope Creep — The natural instinct was to keep adding features. We had to actively lock scope early. Our rule: if it isn't in the user flow, it doesn't exist this weekend.
Real-Time State Without a Database — Tracking live trip state across API calls in-memory required careful design. We keyed a structured Map by trip ID, which stayed clean and fast for the MVP.
Getting the AI Tone Right — Making the companion feel calm and human — not robotic — took several prompt engineering iterations across ElevenLabs and Claude.
Route Observations Without Overclaiming — We deliberately avoided labeling any route "safer," since we don't have real crime data. Writing AI observations that are useful and honest was a real design challenge.
Alert Reliability — Ensuring SMTP fired correctly with the right location and context, even across mid-escalation state changes, required careful backend error handling.

Accomplishments that we're proud of

Shipped a working end-to-end prototype — route selection, live companion conversation, safety monitoring, and trusted contact alerts — all in one weekend
Built a rule-based risk engine that transitions cleanly between Safe, Uncertain, and Risk states without any ML overhead
Integrated a real two-way AI voice companion that feels like a natural conversation, not a chatbot
Practiced disciplined scope management — we cut emotion detection, ML threat scoring, and automatic 911 calling to ship something real instead of something imaginary
Nailed the tone — every word in the UI, from check-in prompts to escalation messages, was written with emotional care

What we learned

Building SafeWalk taught us that restraint is a feature. The hardest decisions weren't technical — they were about what not to build.

We also learned that safety apps live or die on tone. Every prompt, every label, every alert carries emotional weight. Getting that right mattered as much as the code.

Technically, we deepened our understanding of chaining multiple external APIs into a coherent real-time experience, and learned how to model live state machines cleanly without reaching for a database.