Generate synthetic data

Start by specifying the domain, data types (tabular, text, image, time-series), and statistical properties (distributions, correlations, missing rates) of the target real-world dataset. Document the schema including field names, data types, value ranges, and any constraints (e.g., foreign keys, uniqueness). This blueprint ensures the synthetic data will be fit for purpose.

Why Google AppSheet AI: Google AppSheet AI directly supports Natural Language to SQL Schema generation, which is the core need for defining data requirements and schema.

If available, gather a representative sample of real data (anonymized if needed) to extract statistical patterns — distributions, correlations, and anomalies. Perform exploratory data analysis (EDA) using histograms, correlation matrices, and missing value analysis. This step is optional but highly recommended for high-fidelity synthetic data.

Why YData Fabric: YData Fabric provides data profiling, which directly matches the need to analyze a real data sample with tools like pandas-profiling.

Choose a synthetic data generation approach based on your data type and requirements: rule-based (e.g., Faker for simple fields), statistical models (e.g., Gaussian copula for tabular data), or deep learning (e.g., GANs, VAEs for images/text). Configure parameters such as model architecture, training epochs, and privacy settings (e.g., differential privacy epsilon).

Why NVIDIA NeMo Data Designer: NVIDIA NeMo Data Designer is specifically designed for synthetic data generation, matching the need to configure a generation method.

Execute the generation process to produce a dataset of the desired size (e.g., 100,000 rows or 5,000 images). For deep learning methods, train the model on the real data sample (if available) then generate. For rule-based methods, run scripts that produce records conforming to the schema. Save the output in a standard format (CSV, Parquet, JSON).

Why NVIDIA NeMo Data Designer: NVIDIA NeMo Data Designer directly performs synthetic data generation, which is the primary task of this step.

Compare the synthetic dataset against the real data profile (or schema constraints) using statistical tests (e.g., Kolmogorov-Smirnov for distributions, correlation difference). Check for missing patterns, duplicate rows, and edge case coverage. If quality metrics are not met, adjust generation parameters (e.g., increase training epochs, add constraints) and regenerate.

Why NVIDIA NeMo Data Designer: NVIDIA NeMo Data Designer includes model evaluation, which can be used to validate synthetic data quality.

If the synthetic data was derived from real data and needs to be shared publicly, apply additional privacy techniques such as differential privacy, k-anonymity, or field-level masking (e.g., hashing IDs, rounding values). This step is optional but critical for compliance with regulations like GDPR or HIPAA.

Why Tonic AI: Tonic AI explicitly offers data masking, which is the core need for applying privacy and masking.

Finalize the synthetic dataset by adding metadata (generation date, method, schema version, quality report). Package into a deliverable format (e.g., zip file with CSV and PDF report, or API endpoint). Upload to a shared location (S3 bucket, internal data portal) and notify stakeholders with usage documentation.

Why dbt Cloud (AI-Powered): dbt Cloud (AI-Powered) provides AI-generated documentation, which is a key part of packaging and delivering synthetic data.