City crime feeds
Fifteen city-level public safety datasets collected across twelve years, each with its own schema, timestamps, location quality, and offense taxonomy.
- CAD / RMS exports
- Open-data portals
- City-specific schemas
Production architecture
CRIMESENSE / NATIONAL RISK FORECASTING
Data infrastructure first, forecasting system second.
CrimeSense operationalizes twelve years of crime data across fifteen cities into a live 24-hour national H3 forecast. Under the interface is CrimeNet: a production data system that standardizes event feeds, builds model-ready feature contracts, and serves two-stage spatial risk inference at national scale.
15Cities
12YCrime data span
16.7MAudited events
180M+Training examples
25.56MH3-r9 cells
24HForecast horizon
01
DATA FOUNDATION
Data first. Modeling second.
The production system starts with a long-horizon multi-city crime corpus, then progressively standardizes, audits, spatializes, enriches, and materializes it into model-ready and serving-ready assets. The chart below is the actual story of the system.
Canonical event normalization
Raw source records are ingested, cleaned, schema-aligned, and mapped into a unified offense vocabulary so the system can train nationally.
- Bronze ingestion
- Schema alignment
- Canonical offense mapping
Spatial-temporal event backbone
Every event is resolved into a consistent time basis and geospatial representation, forming the event spine that supports both modeling and auditability.
- Timezone resolution
- H3 indexing
- Event spine + QA
National feature joining
The event backbone is enriched with national context layers that describe place, time, weather, and solar state at the modeling hour.
- OSM + built environment
- ACS socioeconomic
- Weather + solar + lighting
Training + serving contracts
The final system materializes model-ready tables for supervised learning and a serving contract used to build future-state snapshots for inference.
- Training table
- Feature store
- Inference snapshots
INPUT SCALE15 cities · 12 years of crime dataHeterogeneous event feeds standardized into one national modeling system.
TRAINING PRODUCT16.7M audited events → 180M+ training examplesEvent records become spatial-temporal supervised learning instances.
SERVING PRODUCTFeature contracts + future-state snapshotsExactly the same system that builds training assets also powers inference.
02
MODELING SYSTEM
Two models. One national risk field.
Once the data system has produced consistent spatial-temporal features, CrimeNet splits the forecasting problem into expected activity and conditional event type.
MODEL INPUTModel-ready national training assetsCanonical events, aligned features, complete serving contracts, and audited spatial-temporal context.
STAGE 01INTENSITY
λ(x,t)
How much activity?
Estimates expected event intensity for each canonical H3 cell under its spatial, temporal, environmental, and socioeconomic context.
- MODEL
- XGBoost
- OBJECTIVE
- Poisson / point-process
- OUTPUT
- events / cell / hour
STAGE 02MARK
P(mark | x,t)
What kind of event?
Given event context, estimates the conditional probability distribution across all modeled crime subtypes.
- MODEL
- XGBoost
- OBJECTIVE
- multi:softprob
- OUTPUT
- 87-class distribution
MODEL OUTPUTNational H3 risk surfaceA live spatial field of expected activity plus conditional subtype probabilities, refreshed over the forecast horizon.
03
FEATURE SYSTEM
Place context meets hourly state.
Both production models share the same 38-feature serving contract: 25 static place features and 13 values rebuilt or resolved for the inference hour.
17
Built environment
Road density, intersections, buildings, POI mix, road composition, and land-use structure.
8
Socioeconomic context
Population, income, poverty, unemployment, vacancy, tenure, age, and vehicle access.
6
Local temporal state
Local hour and weekday plus cyclical encodings resolved in each cell’s timezone.
3
Forecast weather
Forecast temperature, relative humidity, and an explicit weather-availability signal.
4
Solar + lighting
Solar elevation, azimuth, daylight state, and categorical twilight or night context.
Slow-moving national layers that describe the neighborhood and the built environment.
+
Fast-moving values rebuilt for the forecast hour under local time and future environmental state.
→
04
NATIONAL SPATIAL INFERENCE
One model grid. Six serving resolutions.
The model predicts at H3 resolution 9. Lower resolutions are deterministic render-time aggregations for viewing, not separate coarser models.
CANONICAL MODEL GRID
25.56MH3 r9 cells covering the production inference universe.
CRIMENET INFERENCEr9Primary ML prediction resolution.
VIEWPORT SERVING
r9r8r7r6r5r4
Lower zoom levels are deterministic aggregates of the canonical risk field.05
ROLLING FORECAST ENGINE
24 independently inferred future states.
Every forecast hour receives a complete feature snapshot. The system is performing model inference over future environmental state, not merely interpolating one static map.
LIVEINDEPENDENT HOURLY INFERENCE+24H
LIVE
+6H
+12H
+18H
+24H
Forecast meteorologyTemperature + humidity
Future local timeHour + weekday
Solar geometryElevation + twilight
Static place contextSame place, new hour
POWERED BY CRIMENET
Infrastructure beneath the forecast.
The visible product is a risk map. Underneath it is a national geospatial ML stack that builds feature systems, materializes future-state inputs, executes inference, and serves multi-resolution output to the client.
National feature stores
Forecast materialization
GPU inference
LOD generation
Production serving