Integrations
Technologies and workflows Synthex may connect with, each with an accurate status. No integration is currently implemented and no partnership is implied.
Page type: Explains how the platform is designed to fit together.
Integration status key
- Implemented and Verified
- A working integration exists and has been tested.
- Experimental
- Prototype integration work exists but is not supported.
- Exploratory
- Being researched as a candidate technology; no integration code exists.
- Proposed
- An intended interoperability target; nothing is implemented.
Machine learning
Training stays in your environment, with your framework. Synthex's intended role is to supply datasets in open formats and to evaluate the resulting models.
| Technology | Status | Possible pattern |
|---|---|---|
| PyTorchOpen-source deep learning framework. | Proposed | Load packaged datasets through a standard Dataset class that reads the manifest, images and labels. |
| TensorFlowOpen-source machine learning platform. | Proposed | Read the same files with tf.data pipelines; no framework-specific export is required. |
| Common computer vision workflowsClassification, detection and segmentation training and evaluation scripts. | Proposed | Use open annotation formats (COCO-style, YOLO-style, indexed masks) that existing tooling already reads. |
Dataset formats
Open formats are the main interoperability mechanism. No Synthex exporter is released yet; the formats below are documented targets.
| Technology | Status | Possible pattern |
|---|---|---|
| COCO-style JSONImages, categories and annotations (boxes, polygons, RLE) in one JSON file. | Proposed | Detection and instance segmentation datasets. |
| YOLO-style text labelsOne text file per image: class id and normalised box per line. | Proposed | Lightweight detection datasets. |
| JSON annotation metadataPer-sample metadata: source, scenario parameters, seeds, licence. | Proposed | Provenance and filtering by condition. |
| Segmentation masksIndexed PNG masks whose pixel values index the class list. | Proposed | Semantic segmentation datasets. |
Simulation
Simulators are candidate engines for scenario-based data generation and testing. None is integrated, and no vendor relationship exists.
| Technology | Status | Possible pattern |
|---|---|---|
| CARLAOpen-source autonomous driving simulator (code MIT, assets CC-BY), built on Unreal Engine. | Exploratory | Under exploration for research scenarios: a scenario configuration would be translated into simulator settings and sensor definitions. |
| Unreal EngineReal-time 3D engine. | Proposed | Possible rendering backend for photorealistic scenes, directly or via simulators built on it. |
| UnityReal-time 3D engine. | Proposed | Possible environment for custom scenes such as agricultural fields or warehouses. |
| Other compatible frameworksAny simulator that can be driven by configuration and export sensor data. | Proposed | An adapter boundary keeps scenario definitions independent of a single simulator. |
Infrastructure
Generic infrastructure patterns, not services Synthex operates today.
| Technology | Status | Possible pattern |
|---|---|---|
| Cloud object storageStorage for images, masks, manifests and reports. | Proposed | Datasets as immutable, versioned prefixes referenced by manifests. |
| Containerised processingGeneration, validation and evaluation steps packaged as containers. | Proposed | Reproducible environments with pinned dependencies. |
| Batch compute systemsQueued, asynchronous execution of long-running jobs. | Proposed | See the batch processing concept in Architecture. |
Experiment tracking
Reproducibility depends on recording exactly what was run.
| Technology | Status | Possible pattern |
|---|---|---|
| Versioned experiment manifestsJSON configuration recording model, dataset version, protocol, metrics and seed. | Proposed | See the experiment configuration in the Evaluation Engine guide. |
| Common ML evaluation practicesHeld-out test sets, fixed protocols, per-condition reporting. | Proposed | Applied in documentation examples and the research methodology. |
Interoperability principles
- Open formats first: datasets should be usable with existing tools without a Synthex dependency.
- Explicit manifests: datasets, scenarios and experiments are identified by id and version, not by file paths.
- Customer-controlled environments: training and sensitive data stay where their owners keep them.
- Adapters at the edges: simulator- or framework-specific code stays behind a narrow boundary.