HomeSolution
SOLUTION · DEEP DIVE

Three AI engines in one —
a synthesis platform built for structured data

A hybrid engine combining GAN, Diffusion, and Transformer, four validation gates,
and deployment from on-premise to cloud — take a detailed look inside RealDataEcho.

HYBRID ENGINE

What one model can't do,
three engines divide and conquer.

Distributions, fine-grained patterns, and time-series relationships — the three hard problems of structured data synthesis
are each assigned to the best-performing generative model, and a gating layer automatically selects the optimal combination.

CORE ARCHITECTURE

RealDataEcho
Hybrid Engine

Compared with single-model synthesis, distribution similarity improves by +18% and time-series fidelity by +24%. Domain-specific weights are tuned automatically.

3-engine unified gating →
01 · GAN

Distribution replication

Generator–discriminator adversarial training faithfully reproduces the complex joint distribution of the source data.

High-dimensional categorical variables Outlier pattern recovery Sparse distribution preservation
02 · DIFFUSION

Fine-pattern generation

The noise-to-data reverse process restores subtle correlation structures and inter-column dependencies without loss.

Inter-column correlation 0.91+ Precise continuous-variable fidelity Mixed (numeric + categorical) handling
03 · TRANSFORMER

Time series & sequences

Self-attention learns temporal dependencies and long-range patterns to naturally synthesize transaction, log, and event data.

Long-range dependency preservation Automatic seasonality detection Multivariate sequence support
3+
AI ENGINES
0.94
DISTRIBUTION SIMILARITY
+24%
TIME-SERIES ACCURACY
Auto
ENGINE GATING
DATA TYPE COVERAGE

Every structured data type,
handled by a single engine.

The eight column types you meet in finance, healthcare, and marketing — automatically recognized and routed within a single pipeline.

SUPPORTED

Numeric (continuous)

Continuous variables such as amounts, durations, and scores. Preserves distribution shape and even outliers.

SUPPORTED

Categorical

Discrete variables such as regions, grades, and codes. Handles high-cardinality categories reliably.

SUPPORTED

Date & time

Temporal variables such as transaction and enrollment dates. Learns seasonality and day-of-week patterns automatically.

SUPPORTED

Time-series sequences

Sequences such as transactions, logs, and sensor data. Preserves multivariate long-range dependencies.

SUPPORTED

Relational (multi-table)

Multiple tables linked by foreign keys. Synthesizes while maintaining referential integrity.

SUPPORTED

Binary & flags

Churn, consent, and similar flags. Automatically corrects class imbalance.

BETA

Short text

Short strings such as addresses, memos, and product names. Synthesized token by token at the semantic level.

BETA

Geo & coordinates

Latitude/longitude, administrative districts, and more. Anonymizes while preserving spatial distribution.

PRIVACY STACK

Five privacy defense layers
shut down re-identification at the source.

Not a single de-identification technique — a five-stage defense system runs from pre-detection to post-validation. Each layer has its own independent pass criteria, and if even one stage fails, the synthetic output never ships.

Automatic PII detection · NER-based
Identifies and quarantines 18 PII pattern types — national IDs, phone numbers, emails, and more — before training.
Differential privacy (DP, ε ≤ 1.0)
Noise injection during training — a mathematically proven privacy guarantee.
k-anonymity · l-diversity applied together
Protects sensitive attributes and avoids homogeneity, blocking even inference attacks.
Membership inference attack (MIA) defense
Regularization during training so no one can guess whether a source record was included.
1
Sensitive data detection & quarantine
NER-based automatic PII identification · 18 pattern types
PASS
2
Differential privacy noise injection
ε = 1.0 / δ = 1e-5
PASS
3
k-anonymity · l-diversity guarantee
k ≥ 5, l ≥ 3
PASS
4
Synthetic data generation
Hybrid engine output
PASS
5
Post-hoc re-identification risk validation
MIA · distance-based matching · attribute inference tests
0.01%
VALIDATION GATES

Four validation gates
assure quality automatically before release.

Beyond "does it look like real data" to "does a model trained on it perform as well as one trained on the real thing" — measured with objective metrics, and shipped only on a pass.

GATE 01 · STATISTICAL
0.94

Distribution similarity

Measures per-column distribution agreement with the KS test and Wasserstein distance.

GATE 02 · RELATIONAL
0.91

Correlation structure preservation

Distance between Pearson/Spearman correlation matrices — how well inter-column dependencies are preserved.

GATE 03 · UTILITY
0.92

Downstream performance

Model trained on synthetic data vs. model trained on the original — AUC/F1 equivalence testing.

GATE 04 · PRIVACY
0.01%

Re-identification risk

Quantifies source exposure risk with MIA and DCR (distance to closest record).

DEPLOYMENT

Three deployment options
to match your security requirements.

From air-gapped networks to cloud SaaS — on-premise, where your data never leaves, is the most popular choice.

MOST POPULAR

On-Premise

Installed directly on your internal servers. Not a single line of data leaves your network.

  • Full support for air-gapped environments
  • Your security policies stay untouched
  • Recommended for finance, healthcare, and public sector
Hybrid

Private Cloud

Isolated deployment in your dedicated VPC. Cloud efficiency and security together.

  • AWS / Azure / NCP supported
  • Runs isolated in a dedicated VPC
  • Elastic GPU resource scaling
SaaS

Managed Cloud

Ready to use as SaaS. For fast validation in PoC, research, and development.

  • Instant API-based integration
  • Usage-based pricing
  • Zero infrastructure overhead

See the synthesis engine for yourself.

The 3-engine hybrid architecture and four validation gates — all in a single 30-minute demo.