Haven

Map Screen View

Inspiration

Over 12% of Canadians report pain, mobility, or flexibility-related disabilities, yet finding truly accessible spaces is still difficult. We were inspired by the gap between accessibility standards on paper and what people can confidently use in real life. We wanted to build a tool that makes accessibility easier for everyday users to understand, compare, and discover, not just architects or code specialists.

What it does

Haven helps users identify and evaluate accessible homes and public spaces.

Users upload floor plans or building documents.
Haven extracts accessibility-related features from the file (e.g., ramps, elevators, signage, routes).
It compares the extracted content against an accessibility rubric using chunked semantic matching and weighted scoring.
It produces an explainable score (with what was found and what was missing).
It can place the result on a map as a pin/heatmap entry so users can explore accessible places across the city.

How we built it

Vite + React frontend for upload flow, map visualization, settings, reports, and final score views
Django + Django REST Framework backend for APIs, upload handling, geocoding, and report retrieval
SQLite database for locations, comparisons, scores, and chunk-level match results
OpenAI embeddings for semantic similarity scoring between extracted content and rubric criteria
Markdown chunking + cosine similarity to compare building reports against housing/commercial accessibility rubrics
Weighted rubric scoring to prioritize important criteria (e.g., tactile markers, signage, ramps, elevators, visual alerts)
We also added geocoding and reverse geocoding so extracted locations can appear on the map and display readable addresses.

Challenges we ran into

Turning messy extracted text into useful, structured accessibility evidence
Designing rubric comparison that is explainable (not just a black-box score)
Tuning similarity thresholds so incomplete plans still get meaningful results
Handling documents with no address or incomplete location data
Keeping frontend and backend flows consistent (upload -> location confirmation -> final score -> map)
Database schema drift / migration issues while evolving tables for scoring and chunk matches
Making results readable in the UI (top criteria shown clearly instead of dumping everything)

Accomplishments that we're proud of

Building an end-to-end workflow from file upload to explainable accessibility score
Creating a rubric-based scoring system with chunk-level semantic matching and custom weights
Showing why a place scored the way it did (criteria met / not met), not just a percentage
Integrating scoring results with a map and pins for discoverability
Supporting both commercial and housing rubrics with different weighting priorities
Improving UX flows for missing-location cases so users can still reach the final score page

What we learned

Chunking documents by headings/sections is much better than embedding an entire file
Vectors are useful for semantic matching, but they need rule-based checks and good thresholds
Accessibility scoring needs explainability to build trust
Frontend/backend integration details (routing, async scoring, polling, DB writes) matter as much as the ML logic
Iterating on UX with real edge cases (missing addresses, private reports, delayed scoring) improves product quality quickly

What's next for Haven

Add stronger authentication and user accounts (saved history, private/public controls)
Replace workaround logic for private reports with a proper is_public flag in the database
Improve address handling with better geocoding validation and manual correction
Add richer rubric breakdowns (category scores, evidence confidence, compliance summaries)