01PREDICT
Ordered states, not shuffled snapshots.
Learn how the same place moves through time, with transitions, horizons and next-state targets kept explicit.
LONGITUDINAL SPATIAL DATA FOR FRONTIER AI
Train models on how the physical world changes.
Shelob collects public spatial records from around the world and keeps every version it sees. For model builders, that means data about how places change over time, from many independent sources, with the source and licence on every record.
SHELOB / MODEL DATAATTRIBUTED + LONGITUDINAL
MODEL TASK
Which logistics sites are moving towards disruption over the next 24 months? What evidence supports the forecast?
SEPARATE SOURCES
Boundaries
Permits
Freight
Drainage
Road access
Flood history
SHELOBPLACE + TIME + RIGHTS
TRAJECTORY OUTPUT
Persistent place IDs
Ordered state sequence
Transition events
Evidence + rights
INDEXED OBSERVATIONS100s of millions
SOURCE RECORDS86,000+
TRAJECTORY OUTPUTSStates + change events
EVERY OBSERVATIONSource + licence record
01 / THE DATA GAP
To reason about a place, a model needs its history.
That requires more than a larger pile of snapshots. It requires persistent identity, ordered observations, cross-source context and evaluation designed to reveal whether a model generalises across both geography and time.
02REASON
Cross-source context around every place.
Connect infrastructure, movement, hazards, activity, imagery-derived features and records that are rarely modelled together.
03GENERALISE
Evaluation across new regions and future windows.
Measure geographic transfer, temporal reasoning and calibration with held-out places, periods and source combinations.
04GOVERN
Evidence and rights that survive the pipeline.
Keep publisher, provenance, observation time, licence status and permitted use attached through training, evaluation and retrieval.
02 / THE DATA LAYER
One place. Many evidence streams. One ordered history.
Spatial data usually arrives as disconnected files, schemas and snapshots. Shelob resolves those records to real places, preserves when each observation was made and connects signals that are rarely available in one training or retrieval layer.
TRACKED LOCATION / US INDUSTRIAL CORRIDORSUCCESSIVE OBSERVATION WINDOWS
DATA FAMILY202220242026TRAJECTORY
Freight activityRISING
DevelopmentRISING
Drainage headroomFALLING
Road resilienceFALLING
WHAT THE JOIN REVEALS
Pressure is compounding before it appears in any one record.
The model can see the direction of travel, inspect the evidence behind it and distinguish an observed fact from an inferred trajectory.
03 / DATA TRAJECTORIES
Sequences of how real places changed.
Shelob turns successive, source-grounded observations into model-ready place trajectories. Every sequence preserves the state at each window, transition events, forecast targets, related signals, uncertainty, evidence and permitted use.
RESOLVED PLACEHouston Ship Channel / logistics site HX-14
PLACE ID PERSISTS ACROSS TIME2022BASELINE
Industrial use and freight activity are stable.
12 observations establish the initial place state.
2024TRANSITION
Development accelerates as drainage margin contracts.
18 new observations expose the first divergence.
2026CURRENT
Flood and access exposure depart from the historic baseline.
24 new observations show the trajectory strengthening.
MODEL-READY OUTPUTSequence, target and evidence.VERSIONED
- SEQUENCE
- Persistent place ID + ordered states
- TARGET
- Next state + forecast horizon
- EVALUATION
- Geographic and temporal holdout
- EVIDENCE
- Sources + provenance + rights
RESOLVED PLACENorthern California / insured property cluster WF-07
PLACE ID PERSISTS ACROSS TIME2021BASELINE
Seasonal fire exposure remains within the portfolio baseline.
15 observations establish vegetation, access and insured density.
2023TRANSITION
Dryness persists as building density and road dependency rise.
21 observations reveal exposure accumulating across sources.
2026CURRENT
Wildfire and evacuation exposure exceed the historic range.
29 observations support a higher-risk portfolio state.
MODEL-READY OUTPUTSequence, target and evidence.VERSIONED
- SEQUENCE
- Persistent place ID + ordered states
- TARGET
- Next state + forecast horizon
- EVALUATION
- Geographic and temporal holdout
- EVIDENCE
- Sources + provenance + rights
RESOLVED PLACEPhoenix metro / healthcare catchment HC-22
PLACE ID PERSISTS ACROSS TIME2022BASELINE
Heat, population and clinic access remain in balance.
14 observations define the catchment and service baseline.
2024TRANSITION
Extreme-heat days and population begin rising together.
20 observations show demand growing faster than local capacity.
2026CURRENT
Heat exposure rises as travel time to care lengthens.
27 observations indicate a widening access gap.
MODEL-READY OUTPUTSequence, target and evidence.VERSIONED
- SEQUENCE
- Persistent place ID + ordered states
- TARGET
- Next state + forecast horizon
- EVALUATION
- Geographic and temporal holdout
- EVIDENCE
- Sources + provenance + rights
RESOLVED PLACEPort of Rotterdam / energy corridor EN-09
PLACE ID PERSISTS ACROSS TIME2022BASELINE
Industrial demand and grid capacity remain closely matched.
17 observations establish the corridor baseline.
2024TRANSITION
Electrification projects and connection requests accelerate.
23 observations expose the first capacity divergence.
2026CURRENT
New energy demand is outpacing network reinforcement.
31 observations reveal where the transition is constrained.
MODEL-READY OUTPUTSequence, target and evidence.VERSIONED
- SEQUENCE
- Persistent place ID + ordered states
- TARGET
- Next state + forecast horizon
- EVALUATION
- Geographic and temporal holdout
- EVIDENCE
- Sources + provenance + rights
01PRE-TRAIN SEQUENCES
Learn persistent place identity and state transitions.
Train on ordered observation windows across land use, infrastructure, hazards, access and activity, all resolved to one persistent place.
02POST-TRAIN FORECASTS
Predict the next state and explain the evidence.
Create supervised tasks for temporal order, direction of change, forecast horizon and the observations supporting a prediction.
03EVALUATE TRANSFER
Hold out entire places, regions and future windows.
Test whether capability transfers across geography and time, with answer keys, evidence sets, hard negatives and explicit uncertainty.
04GROUND LIVE MODELS
Retrieve the history that makes today meaningful.
Give a model the current state, the sequence behind it and the source, recency and licence context required to use it.
MODEL ARTIFACTS
STATE SEQUENCES
·CHANGE EVENTS
·FORECAST TARGETS
·HARD NEGATIVES
·GEO/TIME SPLITS
Discuss trajectory data ↘04 / WHAT CAN BE CONNECTED
Data families that become more valuable together.
Availability and historical depth vary by location. Every delivery is scoped to the coverage, rights and model task agreed with your team.
Place and boundaries
Parcels, administrative areas, planning zones, protected areas and the geometries that define where something is.
Buildings and land use
Footprints, development, industrial activity, density, surface change and how a place is being used.
Infrastructure and utilities
Roads, rail, ports, energy, water, telecommunications and the networks a place depends on.
Movement and access
Freight, mobility, road access, closures, corridors and the changing connections between places.
Hazards and environment
Flood, fire, heat, water, terrain, vegetation and environmental conditions observed over time.
Economic and civic activity
Approvals, business activity, public assets, services, investment and other signals of change.
Imagery and Earth observation
Aerial LiDAR and imagery-derived features where publishers release them, aligned to place and observation time.
Population and services
Population dynamics, health access, public services, amenities and the human context surrounding a place.
ALIGNED MODALITIES
Explore global coverage ↗VECTOR FEATURES
·IMAGERY-DERIVED FEATURES
·TIME SERIES
·EVENT RECORDS
·RELATIONSHIP GRAPHS
·TEXT METADATA
05 / BUILT FOR MODEL WORK
From pre-training to grounded inference.
Start with a defined model task. Shelob shapes the geography, history, fields, rights and delivery format around the way the data will actually be used.
01TRAINING + POST-TRAINING
Teach models how places and systems evolve.
Train on resolved place sequences, transition labels, cross-source relationships and difficult negative examples instead of isolated coordinate pairs.
02EVALUATION
Test spatial and temporal generalisation.
Evaluate place resolution, temporal order, next-state prediction, cross-source synthesis and calibration across held-out geographies and future windows.
03RETRIEVAL + INFERENCE
Ground a model in current and historical evidence.
Retrieve the latest state, the sequence behind it, source conflicts and the provenance and rights required to support a model response.
04WORLD MODELS
Build next-state and multi-horizon prediction tasks.
Use successive observation windows to learn development, exposure, access and infrastructure trajectories while keeping observations separate from inference.
06 / DATA YOU CAN GOVERN
The source stays with the data. So do the rights.
Every observation carries its publisher, source URL, observation date, original attributes, projection details and a viewable licence record. Your pipeline can filter by accepted usage status before data reaches training, evaluation or inference.
- Trace a model claim back to the evidence used
- Separate permissive, restricted and unknown rights
- Preserve provenance through export and transformation
- Define which records a model or workflow may consume
- PLACE
- Resolved geometry + identifiers
- OBSERVED
- Timestamp + observation window
- SOURCE
- Publisher + original URL
- PROVENANCE
- Transforms + lineage
- LICENCE
- Terms + usage status
- CONFIDENCE
- Match + inference confidence
07 / DELIVERY
Fit the data to the system you are building.
Use a one-off corpus, a continuously refreshed feed or grounded retrieval at inference time. The evidence model remains consistent across every route.
Training corpora
Partitioned datasets, longitudinal extracts and task-specific feature collections delivered with schemas, manifests and data cards.
PARQUET · GEOPARQUET · JSONL · CLOUD STORAGEEvaluation packs
Held-out regions, spatial reasoning tasks, evidence bundles and rubrics designed around the capabilities you need to measure.
DATASETS · ANSWER KEYS · EVIDENCE · RUBRICSGrounded API
Resolve places, find changes and retrieve the observations, provenance and licence context required for a model response.
REST · MCP · STRUCTURED JSONOngoing updates
Refresh the agreed geography and data families as new observations arrive so the model is not fixed to one stale snapshot.
VERSIONED RELEASES · CHANGE FEEDS · MONITORING08 / GLOBAL INDEX
Global spatial data. One place to work with it.
Shelob maintains a worldwide location index across countries, regions, cities, corridors and sites. Data depth varies by place; your partnership defines the coverage and observation history needed for the model task.
Inspect published coverage ↗09 / TALK TO THE DATA TEAM
Tell us what your model needs to understand about the world.
We will map the task to available sources, geographic coverage, observation history, rights and delivery architecture. You will get a practical view of what is ready, what needs to be built and how a first data engagement could work.
USEFUL TO BRINGTarget model, capability and stagePriority geographies and forecast horizonsRequired history, modalities and refresh rateTraining, evaluation, world-model or inference use