Early Access

The Deterministic
Synthetic Data Platform

Tabular · Healthcare FHIR · Industrial · Physics — with cryptographically-sealed evidence on every run. One platform that covers what other vendors cover at most one of. Provably correct, reproducible on demand, and trusted by regulated organisations.

Four Pillars · One Platform
95.69% Benchmark Fidelity
Sealed Evidence on Every Run
ICS / SCADA
SOC / SIEM
ERP / Relational
Regulated

What Organisations Use RadMah AI For

Build and test machine learning models without accessing production datasets
Develop and validate AI systems in regulated industries where real data is restricted
Train security analysts and IDS systems on realistic, labelled attack scenarios
Satisfy data sharing requirements between internal teams, external partners, and regulators
Simulate industrial control system environments for testing, training, and research
Generate reproducible reference datasets for regulatory submissions and independent audits
Accelerate software development with realistic data that behaves exactly like production
Create privacy-safe data for cross-border transfer, partner sharing, and third-party analysis

How RadMah AI Works

Define a contract. Run the engine. Receive synthetic data and an evidence bundle that proves its quality.

1

Define Your Data Contract

Specify your constraints, schema, and requirements upfront. The contract governs exactly what the engine produces — no surprises, no unchecked outputs.

  • Schema and relational constraints defined declaratively
  • Feasibility-conditional hard guarantees: zero violations for feasible contracts
  • Deterministic by design — cryptographically sealed, bit-for-bit reproducible
Constraints
Relational rules
Schema
Tables & types
Determinism
Cryptographically sealed
Regimes
Scenarios
Contract K Ready
Constraint Enforcement Engine
Contract-enforced generation
Constraint enforcement
Scenario & regime transitions
Feasibility certificates
2

Run the Engine

The engine enforces your contract through deterministic constraint enforcement. It generates projections that satisfy all specified constraints — with feasibility certificates when contracts are achievable.

  • Zero hard relational violations for feasible contracts
  • Reproducible runs with auditable hashes
  • Scenario and regime-aware synthesis for complex systems
3

Receive Projections + Evidence

Every run produces two outputs: the synthetic data (projections) and a comprehensive evidence bundle. The bundle contains constraint, determinism, utility, and privacy risk reports — so you can audit exactly what was generated and how.

  • Constraint Report: which rules were enforced, which were relaxed
  • Determinism Report: reproducibility hashes and seed state
  • Utility Report: statistical fidelity and distribution quality
  • Privacy Risk Report: measured and reported, not assumed
Constraint Report
Rules enforced & relaxed
Determinism Report
Hashes & seed state
Utility Report
Statistical fidelity
Privacy Risk Report
Measured & reported

Platform Capabilities

Four pillars — Synthetic Data, Healthcare FHIR, Industrial Simulators, and Developer Platform — covering tabular, clinical, OT/ICS, and programmatic workflows under one tenant-isolated platform.

Mock Data

Instant synthetic data from descriptions

Describe what you need in plain English. The AI drafts a typed schema, generates production-realistic records with domain-appropriate values. No training data required. No uploads. No waiting. Industry-leading accuracy on public tabular benchmarks.

Describe data in plain English — instant generation
AI-drafted schema with type inference
Domain-aware value generation

Synthesize

Train on your data, generate faithful replicas

Upload your real CSV. The platform trains our synthesis engine on your data — learning distributions, correlations, and tail behaviour. Generate unlimited high-fidelity synthetic rows that preserve the patterns in your original data. Privacy-safe by construction. Industry-leading accuracy on public tabular benchmarks.

Upload CSV — platform trains automatically
Preserves distributions, correlations, and tail behaviour
Privacy-safe — no real records in output

Virtual SCADA

Realistic OT data without a single real sensor

Generates streaming telemetry from virtual industrial environments — not statistical approximations, but realistic sensor readings from a deep plant-template library covering all 16 CISA critical infrastructure sectors. Output across 6 live OT protocols: Modbus TCP, OPC-UA, BACnet/IP, MQTT, DNP3, and IEC 61850.

Deep plant-template library across all 16 CISA sectors
6 live OT protocols: Modbus TCP, OPC-UA, BACnet/IP, MQTT, DNP3, IEC 61850
Deterministically seeded — same spec, identical telemetry stream

Virtual PLC

Air-gapped OT controller simulation in your VPC

Simulate programmable logic controllers entirely inside your own infrastructure. Ships as an air-gapped Docker image that runs in your VPC — no external connectivity required. Pairs with Virtual SCADA for realistic end-to-end OT environments without touching a physical PLC.

Air-gapped Docker image — runs entirely in your VPC
No external connectivity required
Pairs with Virtual SCADA for end-to-end OT scenarios

ICS Security

The most realistic ICS attack data available outside a live network

Generates labelled cybersecurity telemetry for industrial control system environments — realistic normal operational baseline traffic alongside sophisticated multi-stage attack sequences mapped to MITRE ATT&CK for ICS techniques, with ground-truth classification at the record level. Purpose-built for the data gap no other tool fills.

MITRE ATT&CK ICS attack techniques with ground-truth labels
Multi-stage ICS attack sequences with realistic baseline noise
Compound mode: adversarial injection directly into a live SCADA session
IDS training, SOC validation, and critical infrastructure research

Constrained Synthesis

Data that obeys your business rules

Generate synthetic data with hard constraints that are mathematically guaranteed — monotonicity, sum bounds, referential integrity, rate limits. No violations possible. Designed for financial, clinical, and actuarial data where constraint violations invalidate the entire dataset.

Monotonicity, sum bounds, rate limits enforced
Referential integrity guaranteed across tables
Zero violations — mathematically impossible

Healthcare FHIR

HL7 FHIR R4 bundles with zero PHI

Generate HL7 FHIR R4 bundles for healthcare development, validation, and compliance workflows. Shipped clinical vocabularies, core resource types, and 100% referential integrity by construction. Zero PHI in output — statistically derived, not anonymised.

HL7 FHIR R4 compliant bundles with shipped vocabularies
Eight core resource types with referential integrity
Zero PHI by construction — safe for development and compliance

AI Assistant

Plain-English driver with cost gates

Plain-English interface to every capability. Specify your job in natural language, review the cost estimate, approve execution. Every conversation is transcribed and cryptographically sealed — no hidden actions, no untraceable workflows.

Natural-language job specification
Cost gates before every engine invocation
Sealed transcripts — tamper-evident, auditable

Autonomous Data Scientist

Autonomous multi-step pipelines with human approval gates

The only autonomous data science agent purpose-built for ICS/OT security. State a goal in natural language, approve the plan, and receive a cryptographically sealed, evidence-validated dataset. Evidence-based self-healing when results fall short — every decision BLAKE3-chained into an auditable AI decision trail.

Goal to pipeline to evidence bundle workflow
Human approval gate before any engine runs
BLAKE3-chained auditable AI decision trail
Developer Platform

Build on RadMah AI

Every capability on the platform is accessible from a typed SDK, a versioned REST API, and signed webhooks. No UI-only workflows — everything you can do in the app, you can do in code.

Connectors for warehouses, RDBMS, object storage, and model hubs are signed, tenant-scoped, and operate under the same evidence contract as interactive runs. Full developer documentation, OpenAPI spec, and example notebooks live on the RadMah AI product site.

Python SDK
Python 3.11+, async-first, fully typed. Install, authenticate, generate — three lines.
REST API
OpenAPI 3.1, signed webhooks, per-tenant rate limits. Version-pinned endpoints.
14 Connectors
Warehouses, RDBMS, object storage, HuggingFace Hub. Signed, tenant-scoped, auditable.
Evidence-backed
Every SDK or API job emits the same sealed evidence bundle as interactive runs. Programmatic auditability.

Evidence Bundles

Every generation run produces a structured evidence bundle — a product primitive that proves the quality and integrity of your synthetic data.

Constraint Report

Documents which constraints were enforced, which were relaxed, and the feasibility status of the contract.

Determinism Report

Reproducibility hashes, seed state, and run metadata so you can replay any generation exactly.

Utility Report

Statistical fidelity metrics — distribution quality, correlation preservation, and edge case coverage.

Privacy Risk Report

Privacy measured and reported. No absolute claims — quantified risk assessment with methodology.

Core Capabilities

Constraint Enforcement

Zero hard violations for feasible contracts, enforced by deterministic constraint enforcement.

Evidence Bundles

Every run produces auditable proof: constraint, determinism, utility, and privacy risk reports.

Cryptographic Sealing

Every dataset is cryptographically sealed at the moment of generation. The output is bit-for-bit reproducible — provably identical, independently verifiable.

Scenario & Regime Engine

Scenario and regime-aware synthesis for industrial and complex relational systems, with explicit transitions.

Reproducible Runs

Reproducible runs with auditable hashes and evidence bundles. Replay any generation for QA, validation, or security.

Privacy by Evidence

Privacy measured and reported — evidence over claims. No absolute guarantees; quantified risk assessment instead.

Fidelity vs Integrity vs Truth

Most synthetic data tools optimise for statistical fidelity alone. RadMah AI distinguishes between three qualities — and lets you decide which matters most for your use case.

Fidelity

Statistical resemblance to real-world distributions

Integrity

Relational and constraint correctness across tables

Truth

Auditable evidence that proves what was generated and how

What Makes RadMah AI Different

The synthetic data market is growing. RadMah AI occupies a distinct position — and the distinction matters.

Other tools generate data. RadMah AI generates data and proves it.

Tools that generate data without evidence ask users to trust the output. RadMah AI generates evidence that makes trust unnecessary — the output can be independently verified, and the evidence makes that verification straightforward.

Other tools treat determinism as a feature. RadMah AI treats it as a requirement.

Determinism in other platforms means consistent output. In RadMah AI it means the output is cryptographically proven to be consistent, and the proof is attached to every dataset automatically.

Other tools serve general use cases. RadMah AI is purpose-built for regulated and industrial environments.

The Virtual SCADA, the ICS Security, and Synthesize are not adaptations of general-purpose tools. They are capabilities built specifically for operational technology, industrial cybersecurity, and regulated research.

Other tools require expertise to operate. RadMah AI's AI assistant makes it accessible to any team.

Domain experts can describe what they need in domain language and get verified output without becoming synthetic data specialists. The assistant handles specification, routing, execution, and review.

Interested in RadMah AI?

RadMah AI is live on radmah.ai. Visit the product site, or book a technical briefing with our engineering team.