HerPath AI

Navigation designed for safety. An AI-powered system that analyzes real environmental risk signals and recommends safer walking routes for women and vulnerable commuters.

Comment

Inspiration

Every day, millions of women make small, invisible calculations before stepping outside — is this street well-lit? Is there anyone around? Should I take the longer way? Standard navigation apps like Google Maps optimize for speed. They don't know — and don't care — whether the fastest route passes through a dark underpass at midnight.

We built HerPath AI because that gap felt wrong. Safety shouldn't be a manual calculation. It should be built into the route itself.

What it does

HerPath AI is an AI-powered walking navigation system that scores routes by safety, not just speed.

When a user enters a destination, the system:

  1. Generates multiple real walking routes via OpenRouteService
  2. Queries OpenStreetMap (Overpass API) for environmental risk signals along each path — streetlamp density, open venues, alleys, tunnels, and pedestrian infrastructure
  3. Passes all of that to an AI agent (Goose CLI) which reasons about the data and returns a structured safety recommendation
  4. Displays the routes on an interactive map with safety scores, risk factors, and contextual safety tips

Additional features:

  • Time-aware scoring — routes are weighted differently at night vs. day
  • Live walking mode — GPS tracks your position and trims the route line as you walk
  • Live location sharing — share your route with a trusted contact who sees your real-time position, updated every 5 seconds
  • Route comparison table — side-by-side safety metrics across all analyzed paths
  • Emergency SOS button — always visible on the results page
  • Auto geolocation — detects and reverse-geocodes your position on load

How we built it

Backend: Laravel 11 (PHP 8.3) with SQLite. A single SafetyScoringService orchestrates the full analysis pipeline.

AI layer — 3-tier cascade:

  • Tier 1 — Goose CLI: The primary agent. Laravel spawns it as a subprocess, passes route + OSM data via a YAML recipe, and parses the structured JSON response.
  • Tier 2 — Google Gemini: If Goose fails or times out, we fall back to the Google AI Studio REST API with a 6-model cascade (Flash 2.0 → Flash 1.5 → Pro → etc.) to handle rate limits gracefully.
  • Tier 3 — PHP rule engine: Pure deterministic scoring as a final fallback, using lighting score, pedestrian activity, and time-of-day weights.

Frontend: Laravel Blade templates with Leaflet.js for maps and Alpine.js for interactivity. All assets bundled locally via Vite — no CDN dependencies.

Data sources:

  • OpenRouteService — real walking route geometry
  • Overpass API / OpenStreetMap — environmental safety signals
  • Nominatim — reverse geocoding

Challenges we ran into

Multi-line template literals in Blade scripts. Our layout inlines all page scripts into a single <script type="module"> block. Blade's @push / @stack mechanism strips newlines in ways that break multi-line JavaScript template literals — causing silent SyntaxError: Invalid or unexpected token crashes that only appeared on certain routes. The fix was rewriting every template literal as a single-line string.

Goose CLI subprocess reliability. Spawning a CLI process per HTTP request is inherently fragile — timeouts, missing PATH variables on the web server user, and JSON parsing edge cases all had to be handled. The 3-tier fallback system was the solution.

Mobile map rendering. The full-screen Leaflet map wouldn't render correctly on mobile until invalidateSize() was called after layout paint. We added both setTimeout calls and a ResizeObserver to cover all cases.

Rate limits on Gemini. The free tier hits limits quickly during testing. The 6-model cascade was built specifically to keep the app functional during a live demo without needing a paid key.

Accomplishments that we're proud of

  • A fully working AI safety routing system built end-to-end in a hackathon timeframe
  • Real open data powering every safety score — no fake or hardcoded values
  • A 3-tier AI fallback that keeps the app alive even when the primary agent fails
  • Live location sharing that works without WebSockets — just polling and a UUID-based share URL
  • A clean, mobile-first UI with glass-morphism panels over a full-screen map

What we learned

  • Goose CLI is a powerful agentic tool but requires careful subprocess management in a web context
  • OpenStreetMap's Overpass API is remarkably rich for safety-relevant data — streetlamps, alleys, tunnels, and active venues are all queryable
  • Blade + Vite is a surprisingly capable stack for building interactive map UIs without a JS framework
  • Designing for safety is a different UX problem than designing for speed — the mental model of the user is fundamentally different

What's next for HerPath AI

  • Incident reporting — let users flag unsafe spots that feed back into future route scoring
  • Historical crime data integration — overlay public crime datasets where available
  • Saved routes and profiles — remember frequently walked paths and alert on changes to their safety score
  • Push notifications — alert a trusted contact if the sharer stops moving unexpectedly
  • Offline mode — cache route data for areas with poor connectivity
  • Expand beyond walking — late-night transit routing with safety-scored stop selections

Built With

Share this project:

Updates