Hybrid SDK

Deploy the RadMah AI SDK on your own infrastructure for air-gapped environments, data sovereignty requirements, and offline evidence verification.

Overview

The Hybrid SDK is the Enterprise deployment where generation pipelines and the Virtual PLC runtime execute on customer infrastructure, so generation stays inside the customer environment. Evidence bundle verification is fully self-contained -- no network calls required.

ℹEnterprise Plan Required

Contact sales@radmah.ai for licensing and deployment planning.

Architecture

The Hybrid SDK deploys three components on customer infrastructure:

Component	Packaging	Description
Python SDK + Engine Core	Python package	Full generation layer for licensed products, including Synthesis, contract-bound constraint enforcement, and Virtual SCADA
Cryptographic Core Module	Compiled extension	Cryptographic hashing, canonical cryptographic serialisation, cryptographically-strong seed-reproducible RNG, constraint projection
VPLC Docker	Docker image	Container-isolated Virtual PLC runtime with live OT protocol servers

System Requirements

Requirement	Specification
Operating System	Linux (real-time-capable Linux kernel recommended for VPLC timing guarantees)
Python	3.10 or later
Docker	Required for VPLC runtime
CPU	4+ cores per VPLC instance (max 50 instances)
RAM	8 GB per VPLC instance (512 MB container limit)
GPU (optional)	NVIDIA with CUDA for GPU-accelerated training (g4dn equivalent)
Network	Air-gapped OK -- no outbound connectivity required

VPLC Docker Configuration

Each Virtual PLC instance runs in an isolated Docker container with security hardening:

Read-only rootfs
All capabilities dropped except NET_BIND_SERVICE
CPU pinning via --cpuset-cpus
Memory limit: 512 MB per container
ULIMIT rtprio=50 for real-time scheduling
Protocol port allocation: range 30000-30999, 1-4 ports per instance

✦Health Monitoring

Built-in health monitoring tracks liveness (5s interval, 3-failure threshold), timing jitter (1000-sample buffer with p50/p95/p99 compliance), and memory usage (80% warning, 95% critical).

Offline Evidence Verification

Evidence bundle verification is fully self-contained. The cryptographic core module provides all cryptographic operations locally:

Cryptographic hash computation for every artefact in the bundle
Canonical deserialisation of sealed contract
Seal verification across every artefact
Determinism proof verification (recompute output hash)

Offline verification

from radmah_sdk import RadMahClient

# Point to local instance (no cloud API)
client = RadMahClient(base_url="http://localhost:8000")

# Verify evidence bundle integrity
result = client.artifacts.verify_bundle(path="evidence/")
assert result.seal_valid
assert result.all_artifacts_valid

LLM Configuration

In air-gapped environments, AI Orchestrator and Agentic Data Scientist features require a locally hosted LLM. The LLM Gateway supports OpenAI-compatible endpoints, including:

vLLM: High-throughput serving of open models
ollama: Local model serving
Any OpenAI-compatible API endpoint

Configure the local endpoint via the TenantLLMConfig with provider_type: openai_compatible and a base_url pointing to your local inference server.

⚠No LLM Fallback

If no LLM is configured, deterministic Synthesis and simulation remain available. AI Orchestrator, Agentic Data Scientist, and Fabricate contract authoring fail explicitly until an approved provider is configured; the platform does not substitute generic values or a hidden authoring path.

Data Sovereignty

✦Complete Isolation

In the Hybrid SDK deployment:

Training data stays inside the customer environment
Generated synthetic data is stored locally
Evidence bundles are computed and stored locally
No telemetry or usage data is sent to RadMah AI
License validation can be configured for offline operation

Deployment Support

Enterprise customers receive dedicated deployment support including:

A named technical contact and agreed escalation terms
24/7 support
Custom SLA terms
On-site deployment assistance (available upon request)
Custom plant-template development for specialised industrial domains