MontreWalk

Landing Page
Map In Use

Inspiration

Winter in Montreal can be brutal for anyone trying to navigate the city on foot. Icy sidewalks, snow-blocked curb cuts, construction detours, and cracked pavement create barriers that make the "shortest route" often impossible to follow. We saw friends with mobility aids, parents with strollers, and seniors struggling with routes that Google Maps claimed were perfectly walkable. We realized navigation apps optimize for distance, not accessibility, and that needed to change.

MontreWalk was born from a simple question: What if routing could account for real-world street conditions?

What it does

MontreWalk is an AI-powered accessible routing platform that helps people navigate Montreal safely by analyzing real street conditions.

Key Features:

Smart Route Selection - Generates multiple route options and scores them on accessibility (0-100) using AI analysis of street-level imagery
Community Reporting - Users can photograph accessibility barriers (missing curb cuts, icy patches, construction) and submit them instantly
AI Analysis - Gemini 2.0 Flash analyzes each image to detect curb cuts, surface quality, and obstructions, providing structured accessibility scores
Real-Time Updates - Integrates user reports with routing to avoid known barriers
Accessibility Dashboard - Shows issues in your area and along your route

How it works:

Enter your destination
MontreWalk samples points every 40m along candidate routes
Google Street View imagery is analyzed by Gemini AI for each point
Routes are scored based on accessibility, not just distance
You get the most accessible path with detailed warnings

How we built it

Architecture:

Frontend (React + Leaflet)

Interactive map with route visualization
Camera integration for instant photo reporting
File upload support for flexibility
Real-time obstacle markers with severity indicators

Backend (Python + FastAPI)

Route orchestration combining Google Routes API with accessibility scoring
Image analysis pipeline using Gemini 2.0 Flash
MongoDB caching to minimize API calls and improve performance
Separate report handler service for user submissions

AI Pipeline:

Street View API fetches panoramic views (N/E/S/W) at sampled points
Gemini 2.0 Flash analyzes images for:
- Curb cut presence (present/missing/unclear)
- Surface condition (good/cracked/uneven/poor)
- Obstructions (poles, debris, snow, etc.)
- Overall accessibility score (1-5)
Deterministic scoring algorithm aggregates results with distance-weighting

Key Technologies:

Google Routes API - Multi-route generation
Google Street View API - Panoramic imagery
Gemini 2.0 Flash - Computer vision analysis
MongoDB Atlas - Caching and user reports
FastAPI - High-performance Python backend
React + Leaflet - Interactive mapping UI
Tailwind CSS - Modern, responsive design

Challenges we ran into

1. API Rate Limits & Costs

Street View + Gemini analysis for every point was expensive
Solution: Implemented MongoDB caching—if a point was analyzed before, reuse the score. Added strategic sampling (40m intervals) instead of analyzing every meter.

2. Deterministic Route Scoring

Initial approach was too simple (just averaging scores)
Routes with long bad sections scored the same as short bad sections
Solution: Implemented distance-weighted scoring where time spent in problem areas matters more. Also added proximity weighting for obstacles.

3. Real-Time Image Analysis

Gemini API quota limits caused failures during testing
Solution: Built exponential backoff retry logic into analyzer.py with 5 attempts and 10s base wait time.

4. Base64 Image Storage

Storing images in MongoDB was hitting free tier limits quickly during testing
Solution: Designed the system to support both base64 (for hackathon demo with ~10 images) and cloud storage migration path (for production with Cloudinary/S3).

5. Polyline Decoding

Google Routes API returns encoded polylines, frontend needed lat/lng pairs
Solution: Implemented custom polyline decoder to convert efficiently without external libraries.

6. Confidence-Based Routing

When should we trust local scoring vs. asking Gemini to decide?
Solution: If route scores differ by >10 points, pick locally. Otherwise, delegate to Gemini for nuanced analysis considering weather, hazard severity, and context.

Accomplishments that we're proud of

Built a working end-to-end system - From photo capture to AI analysis to route visualization in under 48 hours
Real AI integration - Not just a UI mockup—Gemini actually analyzes images and provides structured accessibility data
Smart caching strategy - Reduced costs by 90% while maintaining real-time feel through MongoDB caching
Production-ready architecture - Separated concerns (routing service, report handler, frontend) for scalability
Deterministic scoring algorithm - Distance-weighted aggregation ensures fair comparison between routes
Beautiful, accessible UX - Clean design with smooth animations, intuitive camera integration, and clear visual hierarchy
Community-driven approach - Users can contribute data to help others, creating a virtuous cycle

What we learned

Technical Skills:

Integrating multiple Google APIs (Routes, Street View, Places)
Prompt engineering for computer vision with Gemini
Polyline encoding/decoding algorithms
Geographic distance calculations (haversine formula)
MongoDB indexing strategies for geo-queries
React state management for complex map interactions
FastAPI service architecture and CORS handling

Design Lessons:

Accessibility isn't just about compliance—it's about understanding real user needs
Simple scoring (1-5) is more useful than complex metrics
Caching is critical for any system using expensive APIs
User-generated content can supplement automated analysis
Visual feedback (loading states, progress indicators) builds trust

Domain Knowledge:

Curb cuts are THE critical infrastructure for wheelchair users
Surface quality matters as much as route length
Weather dramatically impacts accessibility (ice, snow, flooding)
Temporary obstacles (construction, food trucks) change daily
Montreal's unique challenges (harsh winters, construction season, hilly terrain)

What's next for MontreWalk

Immediate Improvements:

Weather Integration - Real-time snow/ice/rain data to adjust scores dynamically
User Authentication - Track contributions, build reputation systems
Crowd Verification - Multiple reports of same location increase confidence
Offline Mode - Cache routes and scores for areas users frequent

Feature Expansion:

Transit Integration - Analyze metro/bus accessibility (elevators, ramps)
Business Ratings - Show which restaurants/stores have accessible entrances
Voice Navigation - Turn-by-turn directions with accessibility warnings
Apple Watch App - Haptic feedback for upcoming obstacles
Multi-City Expansion - Start with Toronto, NYC, Paris

Technical Evolution:

Migrate to Cloud Storage - Move from MongoDB base64 to Cloudinary/S3 for scalability
Real-Time Obstacle Detection - Partner with city services for live construction data
ML Model Training - Fine-tune on Montreal-specific accessibility patterns
Edge Computing - Process images on-device for privacy and speed
API for Third Parties - Let other apps access our accessibility scores

Community Growth:

Partnerships with Disability Organizations - Get feedback from real users
City Government Collaboration - Share data to improve infrastructure planning
Volunteer Network - Organized photo collection campaigns for comprehensive coverage
Gamification - Badges and rewards for active contributors