SafePath: AI Navigation That Prioritizes Safety

Homepage Navigation
With Voice Assistant

Inspiration

All of us had just moved to Chicago. One night, walking back around 11 PM, we followed the “fastest route” from a navigation app. Later, we found out that stretch had multiple assault incidents that same week. Nothing happened — but it easily could have.

That stuck with us.

Navigation software optimizes for time, distance, and tolls. It never asks, “Maybe don’t take this street alone at midnight?”

Chicago publicly reports tens of thousands of crashes and thousands of crime incidents. The data exists. It’s open. But it isn’t used to help people navigate more safely.

So we built SafePath.

What It Does

SafePath is an AI-powered navigation app that routes users around danger — not just traffic.

It pulls 49,505 crash records and thousands of crime reports from Chicago’s open data portal, maps them onto a street-level H3 hex grid, and finds a path that is actually safer.

On average, users trade just 2–3 extra minutes for up to 40% less risk.

It works across travel modes:

Drivers — weighted toward crash risk
Pedestrians — weighted toward crime risk
Cyclists — balanced between crash and crime exposure

Users simply talk to it naturally:

“Drive me from O’Hare to downtown during rush hour.” “Walk me from Millennium Park to Navy Pier, it’s late and I’m alone.”

The AI extracts origin, destination, time context, travel mode, and safety priority from a single sentence.

How We Built It

Routing Engine

Each street segment is mapped into an H3 hexagonal cell (resolution 9). Each cell receives a risk score (0–100) based on:

Severity-weighted crash density
Type-weighted crime density
Hour-of-day multipliers
Live weather multipliers

Routing is performed using NetworkX with a custom cost function:

cost = travel_time + (beta × risk_score)

The parameter beta adjusts how much the user prioritizes safety. The AI infers this dynamically from user language.

Two routes are returned:

Fastest route
Safest route

AI Integration (Gemini)

Gemini is the intelligence layer of the system:

Gemini 3 Flash — natural language parsing and conversational routing assistant
Gemini 3 Flash — route safety summaries and contextual explanations
Gemini 2.5 Flash TTS — natural voice navigation output

Parsing and conversational generation run in parallel using ThreadPoolExecutor to maintain sub-second responsiveness.

Tech Stack

React 19 + Mapbox GL JS (Frontend)
Flask + NetworkX + H3 (Backend)
GeoPandas + OSMnx (Spatial processing)
Open-Meteo API (Live weather)

Challenges We Ran Into

Dynamic Risk Modeling: Risk shifts significantly by hour (peak crash density can be ~7x higher than low-risk hours).
Balancing Safety vs Speed: Avoiding extreme detours while still meaningfully reducing exposure.
Real-Time Voice Navigation: Implementing a two-tier alert system for advance and urgent turn warnings.
Data Cleaning: Ensuring all 49,505 crash records were fully geocoded and temporally structured.