CRIMESENSESPATIOTEMPORAL INTELLIGENCE
ModelOpen explorer
GEOSPATIAL MACHINE-LEARNING PLATFORM

SPATIOTEMPORAL
INTELLIGENCE.

CrimeSense models how reported crime intensity changes across space and time using a leakage-safe event spine, sampled point-process support, and spatial, temporal, environmental, infrastructural, and socioeconomic context.

Open explorer View architecture
Statistical intensity modeling. Explicit coverage. No claim of certainty.
41.8781° N87.6298° WH3 · R9
PREDICTED INTENSITYλ(x,t)events / cell / hour
SCROLL TO TRACE THE SYSTEM
17M+RAW CRIME
RECORDS INGESTED
15.95MCANONICAL EVENT
SPINE RECORDS
180M+POINT-PROCESS
MODEL EXAMPLES
74,689UNIQUE EVENT-SPINE
H3 CELLS
15SUPPORTED
JURISDICTIONS
01
SYSTEM DEFINITION

More than a prediction model.

CrimeSense is the public analytical system built on CrimeNet's current 15-jurisdiction data and ML pipeline: immutable lineage, canonical geography, leakage-safe features, explicit eligibility, and GPU-backed analytical delivery.

01

DATA ENGINEERING

Municipal schemas become one canonical event spine.

02

FEATURE SYSTEM

Every cell and timestamp receives spatially aligned context.

03

MACHINE LEARNING

Point-process objectives estimate event intensity, not deterministic outcomes.

04

INFERENCE CONTRACT

Coverage is established before an estimate can be displayed.

05

GPU INTERFACE

Analytical geometry flows into WebGL instead of React DOM nodes.

02
END-TO-END ARCHITECTURE

A traceable path from source to surface.

The current system spans immutable S3/Parquet snapshots, a canonical event spine, exposure-weighted integration sampling, point-in-time feature enrichment, XGBoost training, a FastAPI serving boundary, and the GPU Explorer.

01
DATA PLATFORM

Raw sources

Municipal + county records · OSM · weather · ACS · solar

02
DATA PLATFORM

Versioned lake

S3 · Delta / Parquet · immutable snapshot lineage

03
DATA PLATFORM

Canonical spine

Bronze → Silver → leakage-safe Gold event spine

04
DATA PLATFORM

Integration

H3 support sampling · exposure weights · temporal support

05
DATA PLATFORM

Feature contract

Point-in-time H3 · environment · causal event history

06
ML PLATFORM

Training

XGBoost point process · geographic CV · GPU acceleration

07
INTERFACE

Explorer

FastAPI · Next.js · MapLibre · deck.gl GPU surface

Dagster and Polars orchestrate the current immutable-snapshot pipeline. S3 with Delta/Parquet stores versioned data products, while frozen lineage ties event, integration, environmental, and final-model snapshots together. Typed contracts protect the interface.
03
DATA ENGINEERING

Heterogeneous urban data, one analytical contract.

CrimeSense uses a Bronze–Silver–Gold data contract, but the current pipeline is no longer coupled to one managed compute platform. Dagster, Polars, S3, Delta/Parquet, DuckDB, Python, and SQL carry the active snapshot and feature workflow.

BRONZE / SOURCE-FAITHFUL

Ingest with identity intact.

More than 17 million municipal and county crime records have been ingested across the expanded source footprint, with source identity, lineage, and time semantics preserved before canonicalization.

_ingestion_run_id · _ingested_at_utc
SILVER / CANONICAL

Normalize before joining.

Offense mappings, units, timestamps, H3 identifiers, OSM values, and socioeconomic periods are validated and projected into stable schemas.

source schema → canonical event contract
GOLD / MODEL-READY

Build context without leakage.

The current point-process dataset expands observed events with exposure-weighted integration support to more than 180 million model examples, enriched with temporal, weather, lighting, OSM, ACS, and causal history context.

event rows · integration rows · exposure weights
SOURCE LINEAGE
Municipal + county open data
OpenStreetMap / Geofabrik
Open-Meteo archive
Census ACS 5-year
pvlib solar state
04
TEMPORAL CORRECTNESS

The future is not a feature.

Offline scores are meaningless when future state leaks into historical examples. CrimeSense constructs causal history strictly before each model timestamp, freezes support by split, and keeps the 2025+ test partition sealed during training and validation.

PAST / ELIGIBLEMODEL TIME · tFUTURE / FORBIDDEN
Prior incidents
Weather observations
Calendar and solar state
FEATURE TIME ≤ PREDICTION TIME
Future incidents
Future aggregates
Future feature state
05
GEOSPATIAL ENGINE

The city becomes a graph of stable cells.

H3 is an architectural primitive across spatial joins, aggregation, neighborhoods, model observations, inference, and rendering. The Explorer renders backend-selected H3-r4 through H3-r9 LOD cells derived from one canonical r9 inference surface.

CITY BOUNDARYH3 R9 CELLK=1 NEIGHBORHOODλ(x,t)
FEATURE CONTRACT / CURRENTContext at every cell and time.
35

Causal history

Canonical temporal-history columns in the event-spine contract, built strictly from prior events

06

Baseline temporal

Local hour and day-of-week with cyclical encodings

07

Weather + solar

Temperature, humidity, availability, solar geometry, daylight, and lighting state

17

Built environment

POI, road, intersection, building, land-use, and urban-mix densities

08

Socioeconomic

ACS population, income, age, poverty, employment, housing, tenure, and mobility context

06
MACHINE LEARNING

Event intensity as a point process.

The current baseline uses XGBoost with a Poisson point-process objective over exposure-weighted integration samples. The active geographic-CV baseline uses 37 numeric features plus one categorical lighting feature; richer causal-history columns remain available in the final model contract.

38-FEATURE BASELINE INPUT
XGBOOSTPOINT PROCESSPoisson objective · CUDA hist
λ(x,t)events / cell / hour
CHRONOLOGICAL VALIDATIONTEST ACCESS: FALSE
TRAIN2014–2023
VALIDATION2024
TEST2025–2026-07-24

Current geographic-CV runs use deterministic training and validation samples across five held-out geographic folds. These bars show the global split policy; each source is further clipped to its documented temporal support. The test split remains untouched during model selection and is reserved for final evaluation.

07
INFERENCE COVERAGE

Eligibility before estimation.

A geographic coordinate is not automatically a valid model input. CrimeSense establishes required feature state before displaying intensity; the current event-spine audit reports 99.969% modeled coverage while preserving missingness explicitly.

LOCATION REQUESTlat · lon · time
FEATURE COVERAGEHistoryNeighbor historyWeatherLightingOSMSocioeconomic
FULL MODELPARTIAL · ONLY IF DEFINEDUNSUPPORTED
MISSING DATA ≠ ZERONO PREDICTION ≠ ZERO RISKUNSUPPORTED ≠ SAFE
08
GPU ANALYTICAL INTERFACE

Model output enters the rendering pipeline.

MapLibre owns the geographic camera and label stack. deck.gl interleaves H3 geometry beneath those labels, keeping thousands of cells in GPU-backed layers instead of React DOM.

H3 RESPONSEtyped · cancellable · city-scoped
DECK.GLH3HexagonLayer
GPUfill · extrusion · picking
MAPLIBREcamera · streets · labels
CRIMESENSEINFERENCE EXPLORER
SUPPORTED REGIONChicagoMODEL JURISDICTION · H3 R9 SURFACEPREDICTION TIMEAUG 21, 2024 · 17:00 CDT
Chicago
LOWHIGHER
GPU H3 SURFACETEMPORAL EXPLORATIONEXPLICIT COVERAGE
Open CrimeSense Explorer
09
SUPPORTED GEOGRAPHIES

Fifteen jurisdictions, one spatial vocabulary.

CrimeSense normalizes heterogeneous municipal and county records into one event contract while preserving source identity and local time. The current event spine spans 15 jurisdictions and 74,689 unique H3 cells; support still does not imply complete covariate coverage for every cell-time.

01AtlantaAmerica/New_York
02BaltimoreAmerica/New_York
03Chandler, AZAmerica/Phoenix
04ChicagoAmerica/Chicago
05DallasAmerica/Chicago
06DenverAmerica/Denver
07Fort WorthAmerica/Chicago
08Los Angeles County SheriffAmerica/Los_Angeles
09Marin County Sheriff, CAAmerica/Los_Angeles
10Montgomery County, MDAmerica/New_York
11New York CityAmerica/New_York
12San FranciscoAmerica/Los_Angeles
13SeattleAmerica/Los_Angeles
14Sonoma County Sheriff, CAAmerica/Los_Angeles
15Washington, DCAmerica/New_York
10
TECHNOLOGY MATRIX

Every technology has a role.

The inventory below reflects imported dependencies and implemented code paths—not an aspirational logo cloud.

Data platform

Dagster + PolarsAsset orchestration and lazy columnar feature construction
S3 + Delta / ParquetImmutable snapshots, partitioned storage, and reproducible lineage
Spark / DatabricksUsed for earlier distributed ingestion and large-scale feature assembly
DuckDBLocal spatial joins and boundary processing
Python + SQLTransformation, audit, and analytical interfaces

Geospatial + sources

Uber H3Stable cells for joins, support sampling, features, inference, and rendering
OSM + GeofabrikBuilt-environment features
Open-Meteo + pvlibWeather observations and deterministic solar state
Census ACS / TIGERSocioeconomic context and jurisdiction masks

Machine learning

XGBoostCurrent Poisson point-process baseline and geographic-CV experiments
OptunaDistributed hyperparameter search and reproducible study state
PyTorchNeural marked point-process research track
CUDAGPU acceleration for large training and evaluation runs

Serving + interface

FastAPILocal inference API and typed model-serving boundary
Next.js 16 + React 19Application routing and analytical surfaces
MapLibre GL + deck.glSynchronized GPU map and H3 rendering
TanStack QueryCancellable, jurisdiction-scoped server state
Zustand + ZodLocal UI state and runtime contract validation
RESEARCH ARCHITECTURE · NOT PRODUCTION
11
CRIMENET Ω / RESEARCH TRACK

Beyond the boosted-tree baseline.

CrimeNet Ω remains a research track rather than the selected serving baseline. It explores neural marked point-process formulations over the same leakage-safe spatial, temporal, and contextual contracts.

ARCHITECTURENeural marked point-process researchINTENSITYContinuous-time event intensityMARK SPACECanonical offense taxonomySTATUSExperimental · not serving baseline

Neural history, graph structure, Hawkes-style excitation, multiscale state, and reporting-process components remain research directions. They are not presented as properties of the current XGBoost serving baseline.

Ω0
12
ENGINEERING PRINCIPLES

Correctness is part of the product.

CrimeSense's strongest guarantees are the ones that prevent an attractive interface from overstating what the data and model can support.

01

Temporal correctness

Historical predictions can only use information established before their model timestamp.

02

Spatial consistency

Heterogeneous locations are normalized into stable H3 cells and authoritative jurisdiction masks.

03

Reproducibility

Versioned configurations, deterministic samples, run IDs, and artifacts preserve experiment identity.

04

Explicit coverage

Missing feature coverage can never masquerade as zero intensity or a safe region.

05

Scalable computation

Lazy columnar scans, partitioned Delta tables, and GPU rendering keep large work off the DOM.

06

Typed boundaries

Runtime schemas protect the interface from malformed or semantically invalid inference responses.

CRIMESENSE / DATA → CONTEXT → INTENSITY

FROM RAW URBAN DATA
TO SPATIOTEMPORAL INFERENCE.

Inspect the model surface, temporal state, and coverage contract directly.

Open explorer View model