Crucible E-RAN: Deployment & Architecture Guide

Overview

Crucible E-RAN is an AI-native simulation ("world model") of a radio access network, paired with reinforcement-learning optimization policies that reduce energy consumption 10–15% while respecting operator-defined QoS guardrails. It is a pure software solution: no hardware changes, no RAN vendor lock-in (Nokia, Ericsson, Huawei, Open RAN).

Components

Component	Role	Runtime
Ingest service	Collects performance/energy counters (Kafka, SFTP, or vendor OSS API); vendor-specific adapters	CPU
World model service	Calibrated AI-native simulation of the customer network; predicts KPI/energy impact of candidate actions	CPU (GPU optional)
Policy engine	RL-trained optimization policies; produces per-site energy actions	CPU (GPU for on-cluster training)
Guardrail supervisor	Validates every action against QoS/availability constraints before release; automatic rollback on KPI deviation	CPU
Dashboard & API	Savings analytics, recommendations review, configuration	CPU

All state (telemetry history, model checkpoints, action log) lives on PersistentVolumeClaims; services are stateless and horizontally scalable.

Deployment topology

Platform: Red Hat OpenShift 4.14+ (validated versions on our Red Hat Ecosystem Catalog page). Single-project (namespace) install via a Helm chart supplied by Humanis; deployment is performed jointly with Humanis engineers during the engagement.
Modes: connected (default) or air-gapped (mirror the registry; offline model updates).
Data flow: counters flow in from the OSS; in recommendation mode nothing flows back to the network. In closed-loop mode, configuration actions flow to the OSS interface, gated by the guardrail supervisor.
Security posture: images run as arbitrary non-root UID under the default restricted-v2 SCC; no privileged containers, no host mounts, no ports below 1024; TLS for all external interfaces; subscriber data is never collected — only equipment-level counters.

Sizing

Cluster sizing depends on network scale, ingest volume, and whether policy training runs on cluster or is delivered by Humanis. Your Humanis engineer provides a sizing sheet for your network as part of the assessment; recommendation mode runs comfortably on general-purpose CPU nodes, and GPU nodes are only relevant for optional on-cluster training.

High availability

Dashboard/API and ingest replicas ≥2 behind OpenShift Services; policy engine and world model recover from PVC-backed checkpoints. Loss of the E-RAN deployment never impacts the network: in the worst case, sites simply remain in their current (or default) configuration.

Engagement model

Assessment — analyze exported network data; quantify savings potential.
World model — build the calibrated network replica; train policies in simulation.
Pilot — recommendation mode on selected sites; validate measured savings.
Scale — closed-loop rollout across the network, with the option of performance-based pricing.