Changelog

AI-powered safety prediction and risk assessment APIs for location-based threat analysis.

1.0

Daily and Hourly Predictions with Baseline Model

FeatureEnhancement

Added statistical baseline model providing daily/hourly granularity and detailed crime type breakdowns

What's New

  • Daily Predictions: Get expected crime counts for any specific day, not just weekly aggregates

  • Hourly Breakdowns: Understand when crime is most likely throughout the day with peak hour analysis

  • Detailed Crime Types: See breakdown within categories (e.g., within "Theft", what percentage is vehicle break-ins vs shoplifting)

  • Confidence Intervals: Statistical bounds on predictions (e.g., "expect 12 crimes, 95% confidence between 8-17")

Two Complementary Models

The API now provides insights from two models in a single response:

ModelStrengths
Primary (Deep Learning)Weekly patterns, spatial relationships, probability distributions
Baseline (Statistical)Daily/hourly precision, specific crime types, confidence intervals

New Response Fields

Predictions now include:

  • Expected daily counts by crime bucket
  • Peak hours for each crime type with risk multipliers
  • Top 15 specific crime types within each category
  • Time factors (day-of-week effect, seasonal patterns, recent trends)
  • Historical comparisons (vs day-of-week average, monthly average, recent 4-week average)

Use Cases

  • "What's the worst hour for theft in this area?" - Peak hour analysis
  • "Is today above or below normal?" - Historical comparison
  • "What specific crimes should I worry about?" - Detailed type breakdown
  • "How confident are you in this prediction?" - 95% confidence intervals

Performance

All 10 cities now have baseline models that train in under 60 seconds, enabling rapid updates as new data becomes available.

0.9

Crime Volume Forecasting

FeatureEnhancement

Predictions now include expected crime counts alongside probability distributions

What's New

  • Expected Crime Counts: The model now predicts not just "what type of crime" but also "how many crimes" to expect in each area

  • Per-Type Volume Breakdown: Get expected counts broken down by crime category (e.g., "expect 3 thefts, 2 assaults this week")

Before vs After

Before: "This area has 35% theft probability, 25% assault" After: "This area has 35% theft (expect ~3 incidents), 25% assault (expect ~2 incidents), 8.5 total crimes expected"

All Cities Updated

Every city model has been retrained with dual prediction capabilities:

CityDistribution ErrorCount Correlation
London3.54%0.62
Chicago3.99%0.88
NYC4.33%0.77
LA4.44%0.80
Philadelphia4.69%0.87
SF4.79%0.79
Denver4.82%0.62
Toronto4.82%0.79
Boston5.82%0.85
Seattle6.19%0.86

Improvements

  • 8 cities now under 5% error (up from 5)
  • Boston improved from 7.85% to 5.82% error
  • SF improved from 6.26% to 4.79% error
  • Average error across all cities: 4.74%

API Response Updates

All prediction endpoints now include:

  • expected_total: Total crimes expected in the time period
  • expected_counts: Breakdown by crime type
  • count_mae: Model accuracy for volume predictions
0.8

10 Cities Across 3 Countries

FeatureEnterprise

Added Boston, Toronto, and Denver to reach full 10-city milestone

What's New

Three new cities complete our initial 10-city goal:

  • Boston: 80 grid cells covering greater Boston area
  • Toronto: 238 grid cells - first Canadian city!
  • Denver: 263 grid cells across the metro area (achieved #5 ranking!)

International Expansion

The platform now covers 3 countries:

  • United States (8 cities)
  • United Kingdom (London)
  • Canada (Toronto)

Full Coverage Summary

RankCityPrediction Error
1NYC3.73%
2Chicago3.82%
3London4.02%
4Philadelphia4.85%
5Denver4.92%
6LA5.23%
7Toronto5.90%
8SF6.26%
9Seattle6.65%
10Boston7.85%

Platform Statistics

  • 10 cities across 3 countries
  • 2,766 grid cells total coverage
  • 7 million+ crime records processed
  • 5 cities under 5% prediction error
  • 46 API tests all passing
0.7

Massive Expansion: 7 US Cities Now Live

FeatureEnterprise

Added San Francisco, Los Angeles, Seattle, and Philadelphia in a single day

What's New

Four new US cities added to the prediction platform:

  • San Francisco: 30 grid cells covering the SF peninsula
  • Los Angeles: 585 grid cells - our largest coverage area
  • Seattle: 93 grid cells across the metro area
  • Philadelphia: 163 grid cells covering the city

Coverage Summary

CityGrid CellsPrediction Error
NYC4463.73%
Chicago2783.82%
London5904.02%
Philadelphia1634.85%
LA5855.23%
SF306.26%
Seattle936.65%

Total Coverage

  • 7 cities across 2 continents
  • 2,185 grid cells total
  • 6.1 million crime records processed
  • 4 cities under 5% prediction error

Improvements

  • Standardized crime categories work consistently across all US cities
  • Fixed handling of rare crime categories in cities like Seattle
  • Los Angeles model handles largest grid (585 cells) efficiently
0.6

New York City Now Available

FeatureEnterprise

Added comprehensive crime predictions for all five NYC boroughs

What's New

  • New York City Coverage: Full crime prediction support across Manhattan, Brooklyn, Queens, The Bronx, and Staten Island

  • Best-Performing Model: NYC achieves our lowest prediction error yet at 3.73%, thanks to extensive historical data

Coverage Summary

CityGrid CellsPrediction ErrorRecords
NYC4463.73%1.5M
Chicago2783.82%1.2M
London5904.02%800K

NYC Highlights

  • 446 hexagonal grid cells covering all boroughs
  • 104 weeks of prediction data available
  • Sub-4% error across most crime categories
  • Automatic routing works seamlessly with existing integrations

Three Cities, One API

The unified endpoint now serves predictions for three major cities across two continents. Existing integrations continue to work unchanged - just send coordinates and receive predictions.

0.5

Multi-City Support: London Added

FeatureEnterprise

Crime predictions now available for London, UK alongside Chicago with unified auto-routing API

What's New

  • London Coverage: Full crime prediction support for Greater London using Metropolitan Police data

  • Unified API with Auto-Routing: Send coordinates from any supported city - the API automatically detects location and routes to the correct model

  • International Expansion: First cross-Atlantic deployment, proving the system works with different police data formats and crime categorizations

Coverage Details

CityGrid CellsPrediction Accuracy
Chicago2784.32% error
London5904.02% error

How It Works

No changes needed to your integration. Simply send coordinates:

  • Chicago coordinates receive Chicago predictions
  • London coordinates receive London predictions
  • Unsupported locations return a clear "not covered" response

London-Specific Features

  • Adapted to UK crime categorization from Met Police data
  • 590 prediction cells covering Greater London
  • Full probability distributions matching Chicago format
0.4

Global Grid System Migration

EnhancementPerformance

Adopted industry-standard hexagonal grid system for consistent worldwide coverage

What's New

  • Hexagonal Grid System: Migrated from custom square grids to the industry-standard H3 hexagonal system used by major rideshare and logistics companies

  • Globally Unique Cell IDs: Every prediction cell now has a unique identifier that works anywhere on Earth, enabling multi-city expansion

Improvements

  • Prediction error reduced to 4.32% - 4% improvement from the previous 4.50%
  • Training time reduced by 33% - Faster model updates and iterations
  • Cleaner spatial relationships - Each cell has exactly 6 neighbors at consistent distances

Why Hexagons?

Unlike squares where corners are 40% farther from the center than edges, hexagons have uniform neighbor distances. This creates more consistent spatial patterns for the model to learn, resulting in better predictions.

What This Enables

  • Consistent cell coverage across any city globally
  • Standardized cell sizes (~5 km² per cell)
  • Foundation for multi-city expansion
  • Industry-compatible cell identifiers for integrations
0.3

Crime Prediction API Launched

FeatureEnterprise

First public release of the crime prediction API with single location, batch, and route risk endpoints

What's New

  • Single Location Predictions: Send any latitude/longitude in Chicago and receive a complete crime probability distribution with risk level assessment

  • Batch Predictions: Query up to 100 locations in a single request with summary statistics across all locations

  • Route Risk Assessment: Calculate cumulative crime risk along a travel route with segment-by-segment breakdown and customizable risk weighting

New Endpoints

EndpointDescription
Single PredictionGet crime probabilities for any location
Batch PredictionCompare multiple locations at once
Route RiskAssess safety along a planned route
Coverage InfoCheck available geographic bounds

What You Get Back

For each prediction request, you receive:

  • Probability breakdown across 6 crime categories
  • Dominant crime type and confidence level
  • Overall risk classification (Low, Moderate, High, Very High)
  • Grid cell information for the queried location

Use Cases

  • Delivery routing: Identify safer routes for drivers
  • Real estate: Compare neighborhood safety profiles
  • Urban planning: Understand crime distribution patterns
  • Personal safety: Plan safer travel routes

Model Performance

  • 4.50% mean absolute error (new best!)
  • All crime categories balanced below 6% individual error
  • Strong correlation between predicted and actual patterns
0.2

Sub-5% Prediction Error Achieved

EnhancementPerformance

Redesigned crime categories and optimized model weights to achieve industry-leading accuracy

What's New

  • Redesigned Crime Categories: Restructured crime classifications into more balanced and meaningful groups for better predictions

  • Probability Distribution Focus: Shifted focus from simple classification to full probability distributions - now providing the complete picture of crime likelihood

Improvements

  • Overall prediction error reduced to 4.74% - When we say "35% chance of theft," the actual value is typically between 30-40%
  • Improved prediction quality across all crime types
  • Better balance between common and rare crime categories

New Crime Categories

  • Theft & Burglary: All theft-related incidents grouped together
  • Battery & Assault: Physical altercations and violence
  • Vehicle & Property Damage: Auto theft, vandalism, and property crimes
  • Serious Violent: High-severity incidents requiring immediate attention
  • Other: Remaining categories including drug offenses and public order

What This Means For Users

Previously, the system would tell you "this area is high risk for property crime." Now it tells you the complete breakdown: "35% theft risk, 25% assault risk, 20% quiet, 15% vehicle damage, 5% other." This full distribution enables better decision-making.

0.1

Improved Prediction Accuracy and Model Calibration

EnhancementPerformance

Enhanced crime prediction model with better probability calibration and honest geographic boundaries

What's New

  • More Accurate Probability Predictions: The model now outputs well-calibrated probability distributions, meaning when we say "35% chance of property crime," you can trust that number

  • Smarter Handling of Low-Activity Areas: Added intelligent classification for areas with minimal crime activity, preventing false positives in quiet neighborhoods

  • Honest Geographic Boundaries: Refined the prediction coverage to focus only on populated areas, ensuring accuracy metrics reflect real-world performance

Improvements

  • Prediction accuracy improved to 82% on validated land areas
  • Probability calibration error reduced by 80%+ (from 0.55 to 0.09)
  • Distribution quality score improved significantly (Brier score: 0.72 to 0.15)

Technical Highlights

  • Switched from daily to weekly prediction windows for more stable patterns
  • Implemented soft probability labeling for smoother predictions across activity levels