Understanding the Synthex Platform Architecture
The proposed engineering layers connecting scenario design, data generation, annotation, model training and evaluation, and where conceptual boundaries sit.
Page type: Explains how the platform is designed to fit together.
High-level architecture
Five layers, from the interfaces people use down to generic infrastructure concepts.
Experience layer
- Developer documentationImplemented
- Research workflowsproposed
- Future platform interfacesproposed
Orchestration layer
- Workflow configurationproposed
- Job lifecycle conceptsproposed
- Experiment metadataproposed
Engine layer
- Simulation Engineproposed
- Data Engineproposed
- Evaluation Engineproposed
Data layer
- Dataset artefactsproposed
- Annotationsproposed
- Scenario metadataproposed
- Experiment reportsproposed
Infrastructure concepts
- Computeproposed
- Storageproposed
- Versioningproposed
- Monitoringproposed
Workflow relationships
- Simulation produces scenario outputs: sensor data, ground truth and scenario metadata.
- The Data Engine manages dataset preparation and annotations: it checks quality, then versions and packages datasets with manifests.
- Model training may happen in an external customer environment with the customer's own framework and compute. Synthex does not offer model training as a managed service.
- The Evaluation Engine assesses models and reports findings overall and per scenario condition.
- Evaluation informs new scenarios and dataset improvements: failure categories become the next scenario requirements.
- 01Scenario definition
- 02Simulation Engine: scenario outputs
- 03Data Engine: annotations, QA, packaging
- 04Model training (external environment)
- 05Evaluation Engine: metrics and findings
Evaluation findings inform new scenarios and dataset improvements
Batch processing concept
Generation and evaluation are long-running, so a platform would process them asynchronously. One possible lifecycle:
- 01Job submission
- 02Validation
- 03Queue or scheduling
- 04Compute execution
- 05Artefact storage
- 06Quality checking
- 07Completion status
Data governance
- Dataset provenance
- Every sample records whether it is real, synthetic or derived, and how it was produced.
- Metadata
- Scenario parameters, seeds, annotators and versions travel with the data, not in separate notes.
- Access controls
- Datasets are only as open as their licence and governance allow; sensitive data stays where its owner keeps it.
- Versioning
- Datasets, scenarios and models are referenced by immutable versions so results stay comparable.
- Usage rights
- Licences and usage restrictions are recorded in the manifest and respected downstream.
- Traceability
- A report can be traced back to the exact model, dataset version, configuration and seed.
- Research reproducibility
- Anyone with access can re-run an experiment from its recorded configuration.
System boundaries
The proposed design keeps clear boundaries between what Synthex would operate and what stays with the customer:
- Customer environment: model training, model weights, and any sensitive or regulated source data, unless an engagement explicitly agrees otherwise.
- Exchange format: datasets, manifests, scenario configurations, experiment configurations and reports, all in open, documented formats.
- Simulator boundary: scenario definitions stay independent of any single simulator, with adapters translating them.
- Public documentation boundary: this portal never publishes hosts, private endpoints, account identifiers, credentials or operational runbooks.