GRDF: Augmented IT on a Sovereign AI Foundation

CASE STUDY · ENERGY & PUBLIC SERVICE

Augmenting spec, dev, test and ops on a foundation we own.

Adservio tools the four phases of the software lifecycle on a sovereign AI platform, for the operator of the French gas distribution network.

Client

GRDF

Expertise

Spec-Kit · OpenSpec · Coding agents · AI tests · evals · Augmented ops

Tech Stack

Sovereign platform · AI Act · Gemma · gpt-oss · Mistral

Engagement

France · Since May 2025

CONTEXT

Project Context

GRDF operates the French gas distribution network. Sovereignty constraints require data and models to remain within the controlled perimeter, without giving up on an IT function that becomes augmented across the entire software lifecycle, from framing to operations.

Since May 2025, Adservio has been augmenting spec, dev, test and ops on a self-hosted LLM platform, with each validated case capitalised into a growing agent marketplace.

"We are a critical infrastructure operator: every piece of data must stay within our perimeter. We want GenAI that is useful across the whole lifecycle, from framing to operations, on a platform that belongs to us.", IT Strategy Department, GRDF

Strategic Objectives

OBJECTIVES

(01)

Cover the whole SDLC

Augment the four phases of software engineering end to end: spec, dev, test and ops, with no blind spot between framing and operations.

(02)

Strong sovereignty

Keep data and models within the controlled perimeter of the critical infrastructure operator, with no exit outside that perimeter.

(03)

AI Act compliance

Guarantee traceability of decisions and auditability of every GenAI usage, in line with the requirements of the European AI regulation.

(04)

Capitalise pillar by pillar

Build an agent marketplace that grows validated case after validated case, rather than a collection of isolated POCs with no future.

Solutions Delivered by Adservio

SOLUTIONS

Adservio tooled the four phases of the lifecycle: spec-driven upstream, coding agents under harness in dev, adversarial tests and evals in quality, augmented runbooks in ops. All on a self-hosted LLM platform (Gemma, gpt-oss, Mistral) backed by Anthropic directly as a vendor. The agent marketplace capitalises every validated case, pillar by pillar.

(01)

Spec

Spec-driven upstream with Spec-Kit and OpenSpec: living specs that remain the product's source of truth, from framing to delivery.

(02)

Dev

Coding agents under harness with rigorous context engineering, to accelerate development on validated recurring patterns.

(03)

Test

Generated tests, adversarial evals and regression guardrails, to raise coverage without adding to the quality teams' workload.

(04)

Ops

Augmented runbooks, assisted RCA and self-healing, to reduce resolution time and stabilise day-to-day operations.

(05)

Sovereign platform

A self-hosted LLM platform (Gemma, gpt-oss, Mistral) backed by Anthropic directly as a vendor: models and data stay within the perimeter.

(06)

Governance

Strategy committees (IT Department · business owners · delivery owners), evangelist circles and domain referents: GenAI spreads in a steered and durable way.

Results

MEASURED FIGURES

−30%

Engineering cycles

Dev, on validated recurring patterns. Coding agents accelerate delivery where the patterns are proven and capitalised.

+50%

Test coverage

Automated generation and adversarial evals: coverage progresses without adding to the quality teams' workload.

100%

Data sovereignty

No exit outside the perimeter. Models and data stay under the control of the critical infrastructure operator.

40+

Agents in the catalogue

Capitalised pillar by pillar: Spec, Dev, Test, Ops. The marketplace grows validated case after validated case.

Impact

WHAT IT CHANGES

Accelerated cycles

Coding agents under harness cut engineering cycles by 30% on validated recurring patterns, where the schemes are proven and capitalised.

Reinforced quality

Automated test generation and adversarial evals raise coverage by 50%, without adding to the quality teams' workload.

Sovereignty held

No exit outside the perimeter: models and data stay under the control of the critical infrastructure operator, on a self-hosted LLM platform.

Durable capitalisation

A marketplace of more than 40 agents capitalised pillar by pillar (Spec, Dev, Test, Ops), growing validated case after validated case. It is the IT function that learns, not a supplier.

“The 4 pillars deliver at pace, every capitalised agent enriches the marketplace, the foundation stays sovereign. It is the IT function that learns, not a supplier.”

Head of AI Innovation

GRDF