RadMah, Inc. joins NVIDIA Inception to scale deterministic synthetic data and ICS/OT simulation
RadMah AI's move to larger runs gains momentum with acceptance into NVIDIA Inception's ecosystem for AI startups.
London, UK — 15 September 2026 — RadMah, Inc., creator of RadMah AI, has been accepted into the NVIDIA Inception program. NVIDIA approved the application on 14 September 2026. RadMah builds deterministic synthetic data and process-model-driven ICS/OT simulation for sectors where the output has to survive an audit — healthcare, energy, finance and industrial control.
Data you can prove, not just generate
RadMah AI generates tabular, time-series and HL7 FHIR R4 data. Synthesize runs RadMah's Differentiable Voronoi Flow Matching engine; there is no fallback model. On the public mostlyai-qa benchmark it scored 95.69% (Adult 24K, 200 epochs) with 99.28% univariate accuracy — a representative result on a public dataset, not a guarantee for any specific one.
Every job ships a signed .tar.zst evidence bundle: the sealed Contract K job spec, run log, quality report, privacy analysis, lineage, relation-closure certificate, engine SBOM, a BLAKE3 hash chain and a determinism proof. It verifies offline with RadMah's verifier tool, included in the RadMah SDK. The same Contract K and the same seed produce bit-for-bit identical output, on any cluster, on any later date.
Plant, simulated at the wire
The industrial side runs Virtual SCADA and Virtual PLC, streaming wire-level telemetry over industrial physics models, across 67 plant templates. RadMah AI supports six OT protocols (Modbus TCP, OPC-UA, BACnet/IP, MQTT, DNP3, IEC 61850); DNP3 and IEC 61850 live protocol servers are Enterprise VPLC. The result is cyber-range output: labelled datasets for IDS training, incident rehearsal and control-system testing that no operator can safely produce from a running plant.
Where the headroom goes
RadMah plans to evaluate accelerated compute along three paths. The first is training and sampling in the tabular core, where longer epoch budgets and wider hyperparameter sweeps inside the same wall-clock are how fidelity on long-tailed data is won. The second is the quality battery that gates every job — statistical tests, constraint satisfaction, relation closure, per-column drift — which sets how quickly an engineer gets a verdict on a run. The third is the concurrency ceiling on simulated sites, which decides how large an ICS dataset can get.
One constraint travels with all of it. Any faster execution path has to return the same bytes under the same contract and seed, or it does not reach a customer job. That is the product, and it is not up for trade.
What comes next
RadMah Sentinel, the company's runtime control and evidence layer for AI agents, works on the same principle: decisions an operator can inspect afterwards, on the record. RadMah will report what membership changes as it lands, at radmah.ai. Teams scoping a regulated programme can request a synthetic-data evaluation at itlox.com/contact — and verify the bundle themselves.
Joining NVIDIA Inception is a milestone for RadMah, and a practical one. Synthetic data is easy to generate and hard to prove — every RadMah AI job comes back with a signed evidence bundle and a determinism proof, verifiable offline on the customer's own machine. Membership gives us room to push the engine harder while that guarantee holds.
Hospitals, banks and industrial operators cannot hand us their production data, and they will not accept a synthetic substitute they are unable to audit. RadMah AI answers both at once. Membership gives our engineers training and ecosystem access at the point in our growth where it counts, and it arrives as verifiable synthetic data becomes a procurement requirement rather than a preference.
Where membership goes to work
Compute headroom where it counts
Preferred pricing on select NVIDIA hardware and software, applied to capacity RadMah already budgets for. Cloud credits from NVIDIA and its partners are available to request, and the team plans to evaluate them for longer epoch budgets and wider hyperparameter sweeps on the tabular core.
Bigger industrial runs
Longer simulated horizons and more concurrent simulated sites across the supported OT protocols. Cyber-range and IDS datasets are limited by how much plant you can simulate at once, and RadMah plans to raise that ceiling.
A deeper engineering bench
Free self-paced NVIDIA Deep Learning Institute courses and discounted expert-led workshops, aimed first at the engineers maintaining the generation, evaluation and simulation paths.
Current platforms and ecosystem reach
Access to the latest SDKs, model libraries and developer platforms, with full access to the NVIDIA Developer Forums — plus technology guidance and connections within NVIDIA's ecosystem, and go-to-market and visibility opportunities that widen the routes to the regulated teams RadMah builds for.
Next step
Run it on your data, and verify the bundle yourself.
About ITLOX
ITLOX builds high-assurance software for regulated teams that need privacy-safe data, secure communications and audit-ready evidence. RadMah, Inc. (Delaware, USA) is an ITLOX company and operates RadMah AI — deterministic synthetic data across tabular, time-series and HL7 FHIR R4 output, with process-model-driven ICS/OT simulation and an evidence bundle on every job — and RadMah Sentinel, runtime control and evidence for AI agents. Sibling company AegisWire, Inc. builds post-quantum secure communications, and MahCare AI is the group's healthcare operations platform. ITLOX LTD (England and Wales, no. 16587309) is the UK service company, and the founding team is based in London. Learn more at itlox.com and radmah.ai, or request an evaluation at itlox.com/contact.
Trademarks
NVIDIA, the NVIDIA logo and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. RadMah, Inc. is a member of the NVIDIA Inception program; membership does not constitute an endorsement, partnership or investment by NVIDIA.