Inference Public

Insights

The decisions behind
reliable engineering.

Practical notes on infrastructure, software and the discipline of delivery.

Engineering perspective

RAG Is a Data Product: Building Trustworthy Retrieval on AWS

For public-sector RAG, document authority, permissions, withdrawal and provenance determine whether generated answers can be trusted. How to engineer those controls on AWS.

Engineering guidance

Read the article

Engineering perspective

Machinery of Government Changes and Your AWS Organization

When functions move between departments, AWS accounts, keys, commitments, access and logs have to change hands too. What AWS now allows, what it leaves to you, and how to show the receiving department can run the service.

Engineering guidance

Read the article

Independent analysis

What Published Public-Sector AI Projects Actually Show

What published public-sector AI projects show about documents, casework and access to services, and what the evidence has yet to establish.

Research and analysis

Read the article

Software engineering

What a casework service must remember after a retry

A timeout tells the caller that it did not receive an answer. It does not establish whether the service committed the work.

Read the article

FinOps

Who is responsible for the part of the AWS bill nobody recognises?

Shared networking, logging and security costs do not become somebody else's problem because they sit outside an application account.

Read the article

AI engineering

Putting a generative AI feature into service

A usable AI feature needs a defined task, an evaluation that can reject it and an operating arrangement for the answers it should not give.

Read the article

Reliability

Backup and recovery on AWS: an introduction

Recovery objectives, AWS Backup, locked vaults and why a backup proves little until it has been restored.

Read the article

AI engineering

Verifying AI-generated code and tests

A passing test is useful when it would object to the wrong behaviour. Start by deciding what wrong looks like, independently of the implementation.

Read the article

Cloud foundations

Where Control Tower fits in a cloud operating model

What Control Tower automates, what its controls cover, and the ownership decisions it leaves with you.

Read the article

Platform engineering

Your build pipeline is a production system.

Why the system that deploys production should be owned, secured and recoverable in the same way.

Read the article

Delivery discipline

What an embedded DevOps engineer needs to succeed

Mandate, access, ownership and cooperation with operations for an engineer joining your team.

Read the article

Transformation

What is still running when the programme closes?

Closure criteria, decommissioning, funding the running costs and keeping knowledge after the delivery team leaves.

Read the article

Cloud foundations

Planning an OU strategy that outlasts the next reorganisation

Grouping accounts by the rules they need, the limits to design within, and the procedural OUs worth planning early.

Read the article

Transformation

Modernising an AWS estate for cost control

Visibility, ownership, low-risk savings, commitments and the modernisation that changes the cost curve.

Read the article

Cloud foundations

The trouble with VPC peering at scale

What peering does and does not do, why a full mesh grows so quickly, and how to plan the way out.

Read the article

Platform engineering

What does your team still own on Amazon EKS?

Upgrades, access and the responsibility split once AWS runs the control plane.

Read the article

Delivery discipline

A deployment pipeline that produces its own evidence.

Every release should leave a trail that shows what changed, who approved it, how it was verified and how it can be undone.

Read the article

Platform engineering

Argo CD and the record of what is running.

How GitOps changes release approval, drift handling and rollback.

Read the article

Cloud foundations

A landing zone needs owners as well as accounts.

A new account is useful when a team can deploy, investigate a fault and get a decision. The landing zone has to make all three possible.

Read the article

Delivery discipline

What should operational acceptance establish?

An acceptance meeting should establish what the receiving team can operate, which evidence supports that judgement and what remains unresolved.

Read the article

Reliability

What does a recovery test actually demonstrate?

A restored database is one result. A usable service is another. Choose the finish line, record the conditions and keep the limits beside the result.

Read the article

21 articles

Start a conversation

Let’s talk about what you need to build.

Bring the service, the constraint or the engineering question. We’ll work through a sensible starting point.

Discuss a project