Engineering the backbone for national-scale energy AI
This national-scale AI energy client operates one of the world’s most data-intensive exploration environments. Through its AI arm, Inception, they launched a program to convert decades of geological research into real-time predictive intelligence for drilling strategies.
This was not a standard cloud deployment. It required an AI platform capability situated within a critical national infrastructure environment characterised by:
- Operating within a highly restricted network access model.
- Workloads measured in petabytes, processing high-fidelity subsurface data.
- Centralised control over cloud and data platforms across multiple distinct business units.
Challenge
The objective was to move from experimental AI to industrial-grade operations.
The platform needed to support H100 GPU clusters for training and inference while serving two distinct outputs: chat-based interactions and advanced 3D visualisation from a single backend.
The constraints were absolute:
- Subscriptions had to be air-locked, reachable only from the customer network
- The architecture had to fit a rigid multi-account model without stifling developer velocity
- Meeting strict internal security controls while processing petabytes of research data at speed
Solution:
A governance-first DevOps architecture
We embedded within the platform engineering function to deliver an environment that was repeatable, auditable, and fast to extend.
1. Tech stack
The engineering core relied on Kubernetes (K8s) and ArgoCD for GitOps-driven delivery, managed via Terraform. On the data layer, the platform leveraged Azure AI and Azure ML workspaces integrated with massive Data Lakes.
2. Infrastructure as Code (IaC) framework
To eliminate configuration drift, a modular IaC framework replaced manual provisioning. This ensured that every environment spun up automatically inherited the required security, networking, and governance policies.
3. "Air-lock" strategy
Deployment patterns were architected to strictly respect the air-gapped nature of the subscriptions. This approach maintained high-throughput processing pipelines for ML and data workloads without violating the requirement that resources remain accessible only via the customer network.
Outcome
The result was a repeatable, shared internal capability rather than a single-use product.
- New AI use cases can now land on the platform without re-engineering the core infrastructure
- Every change to the environment is tracked via GitOps, satisfying regulatory requirements
- Subsurface teams moved from manual data handling to real-time AI guidance, powered by a platform that abstracts the complexity of the underlying petabyte-scale compute
1PB+
of subsurface data processed in real time across H100 GPU clusters
100% air-locked
subscriptions accessible only via the customer network
0 manual steps
every environment auto-inherits security, networking, and governance policies
1 platform
new AI workloads land without re-engineering the core infrastructure
