Skip to content

Synthex Data Engine

Synthetic Data, Built for Real-World AI

Develop scenario-driven datasets with structured annotations, consistent metadata and quality-focused workflows.

Engine 01Platform VisionIn development — not generally available.
Dataset pipelineIllustrative

01Source scenario

scenario: urban_crossing
conditions: [fog, dusk, rain]
targets: [pedestrian, cyclist]

02Generated samples

03Annotation

04Validation

  • Schema valid
  • Labels complete
  • Balance reviewed
  • Human review

05Dataset package

dataset/v0.1
Illustrative Data Engine pipeline: a source scenario produces generated samples, which are annotated, validated and packaged as a versioned dataset.

Core capabilities

From generated samples to dataset packages

Six planned capabilities cover the path from a scenario description to a reviewed, versioned dataset. None of them is generally available yet.

  • Platform Vision

    Synthetic Data Generation

    Produce images and sensor-like outputs from scenario configurations with controlled variation.

    Purpose
    Cover conditions and classes that are rare, costly or sensitive to collect.
    Expected output
    Generated samples with the parameters that produced them.
  • Platform Vision

    Automated Annotation

    Derive labels from the generation process itself, where ground truth is known by construction.

    Purpose
    Reduce manual labelling for boxes, masks and class labels.
    Expected output
    Machine-generated labels linked to each sample.
  • Platform Vision

    Human-in-the-Loop Review

    Route uncertain or high-impact samples to reviewers with clear acceptance criteria.

    Purpose
    Keep people in control where automated labels are not sufficient.
    Expected output
    Reviewed labels with reviewer decisions recorded.
  • Platform Vision

    Dataset Curation

    Select, balance and split samples against coverage goals rather than convenience.

    Purpose
    Build datasets that reflect the scenarios a model must handle.
    Expected output
    Curated train, validation and test splits.
  • Platform Vision

    Quality Assurance

    Run consistency, completeness and distribution checks before data is released.

    Purpose
    Catch label errors, gaps and imbalance early.
    Expected output
    Quality review summary attached to the dataset version.
  • Platform Vision

    Dataset Packaging and Export

    Package data, labels and metadata in documented, versioned structures.

    Purpose
    Hand datasets to training pipelines without format wrangling.
    Expected output
    A versioned dataset package with manifest and metadata.

How to read status labels

Platform Vision
A planned capability that defines where the platform is heading. Not yet available.
Research Direction
Active exploration of methods and questions. No prototype or results yet; not a product.
Illustrative Demonstration
A visual or interactive explanation using sample content. No live system or real results behind it.
Validated Implementation
Implemented and verified against defined criteria. No capability on this site is presented at this level yet.

Dataset workflow

Every step leaves a traceable artefact

Select a stage to see what goes in, what comes out and what is recorded. The workflow describes the intended design of the Data Engine.

Stage 01 of 07

Input Requirements

Teams describe the task, target classes, conditions of interest and quality expectations.

Inputs

  • Task definition
  • Class list
  • Coverage goals

Recorded outputs

  • Dataset requirements record

Dataset types

Supported and proposed data representations

Representations the Data Engine is designed around. They describe the intended scope, not formats that are available today.

Supported or proposed

Representations describe the intended scope of the Data Engine, not formats available today.
  • RGB images

    Rendered or captured colour images, the base modality for most vision tasks.

  • Segmentation masks

    Pixel-level class or instance labels aligned to each image.

  • Bounding boxes

    Object locations and classes for detection tasks.

  • Classification labels

    Image- or region-level categories with optional attributes.

  • Medical imaging research datasets

    Research-oriented imaging volumes and lesion annotations. Research use only.

  • Sensor metadata

    Camera parameters, poses, timing and scene conditions recorded per sample.

Healthcare example

An illustrative medical imaging research workflow

How the Data Engine could support lung lesion segmentation research on CT imaging. This is a research direction; it would use public, licensed research collections such as LIDC-IDRI.

Research Direction
  1. 01

    Research CT Scan

    Start from de-identified research imaging obtained under its licence terms.

  2. 02

    Preprocessing

    Resample, normalise intensity and select regions of interest consistently.

  3. 03

    Lesion Segmentation

    Reference lesion masks from expert annotations define the learning target.

  4. 04

    Synthetic Data Augmentation

    Generate controlled variations to supplement scarce or imbalanced examples.

  5. 05

    Annotation and Quality Checking

    Check that masks remain anatomically plausible and consistent after augmentation.

  6. 06

    Research Model Evaluation

    Compare models on held-out data with transparent, task-appropriate metrics.

Research context, not clinical evidence

Research direction only. Synthex does not own clinical datasets, has no clinical validation or regulatory approval, and reports no medical performance metrics here. Not for clinical use. See the Healthcare & Life Sciences solution.

Export concept

A dataset package you can inspect

Explore an illustrative package layout. Select a file to preview its sample contents. This is a static illustration — no export service is connected.

dataset/

Illustration · not a live export

dataset/manifest.json

Package version, splits, formats and checksums.

{
  "name": "example-urban-crossing",
  "version": "0.1.0",
  "status": "illustrative-sample",
  "splits": { "train": "…", "val": "…", "test": "…" },
  "modalities": ["rgb", "segmentation_mask"],
  "annotation_format": "coco-style-json",
  "generation_config": "configs/urban_crossing.yaml",
  "checksums": "sha256 per file"
}

Proposed package conventions

Platform Vision
Layout
images/ · masks/ · annotations/ · metadata.json · manifest.json
Annotations
COCO-style JSON for boxes and instances
Masks
PNG, one class index per pixel
Metadata
Per-sample scenario parameters, camera and review status
Versioning
Semantic version + per-file SHA-256 checksums

Conventions under design. Formats may change before any release.

Next steps

Build Better Datasets for Your AI Research

Tell us about the dataset you need, the conditions that matter and how you measure quality. We are open to scoped research and pilot discussions.