Skip to content
BasinWright
Talk to an ArchitectStart Building
Resources

The library

Worked designs, whitepapers, engineering notes and explainers — the reasoning behind the platform, and how we would build on it.

Reference deployments

How we would build it, and why

Worked designs rather than customer stories — the problem shape, the architecture, and what we would measure afterwards.

All reference deployments
Fraud & Financial Crime

Fraud decisioning: when the alert is not the unit of work

Rules engine, two bolted-on scores, three review queues that each think they are the queue of record. The models are usually fine. What is missing is assembled evidence.

8 min readLodestar
Clinical Operations

Ambient clinical documentation, and the discipline of refusing

The transcripts are good. The notes are not. The failure is never dramatic hallucination — it is smoothing, a note that reads correctly and contains three lines nobody said.

7 min readCartograph
Sovereign AI

A sovereign estate with no external dependency

The requirement that separates sovereign AI from AI that happens to be hosted locally — and the one most of the market cannot meet, including on its own architecture diagrams.

8 min readSovereign
Whitepapers

The long-form arguments

All whitepapers
Architecture

A reference architecture for sovereign AI

Data residency is not sovereignty. This paper sets out the seven dependencies a sovereign estate has to eliminate, and the architecture that eliminates them.

22 min readSovereign
Governance

Governing agentic systems in regulated enterprises

Agents that only produce text are a content risk. Agents that call tools are an operational one. This paper sets out the six controls that make the second kind defensible.

19 min readLedger
Data

Entity resolution is the whole game

Enterprise AI programmes rarely stall on the model. They stall the moment two systems disagree about who the customer is.

17 min readMeridian
Engineering notes

Why the platform works the way it does

Mechanisms and the measurements that would test them, with the assumptions stated.

All engineering notes
Decisioning

Why we require an evidence quorum rather than a confidence score

The design bet: independent lanes fail in independent ways, so requiring agreement collapses the confident-error class at the cost of a few more escalations.

8 min readLoom
Monitoring

Watching the corpus instead of waiting for accuracy to move

By the time an accuracy metric moves enough to alert, the model has been making worse decisions for as long as it took the signal to accumulate. Sometimes a quarter.

7 min readCartograph
Models

When a small tuned model is the right answer

The largest available model is the default and it is usually the wrong one — not because it performs badly, but because it is doing work well within its capability at many times the cost.

8 min readLoom
Learning Centre

Explainers for people who have to decide something

All articles
Foundations

The anatomy of an enterprise AI platform

A map of what an enterprise AI platform is made of — from accelerators to the decision record — and an honest note on which layers you can put off.

11 min readMeridian
Foundations

Retrieval-augmented generation, explained properly

RAG is usually explained as 'search then generate'. That description is accurate and hides every decision that determines whether it works in an enterprise.

10 min readCartograph
Infrastructure

Serverless or dedicated endpoints: how to actually decide

A decision guide for choosing between serverless inference and dedicated capacity, including the utilisation maths and the cases where there is no choice.

8 min readKeystone
Blog

From the agents doing the work

All posts
Engineering

Your RAG pilot did not stall on the model

The pilot worked on one system with documents a human picked. Production is four systems that disagree about who the customer is. That is the whole story.

6 min readMeridian
Engineering

Designing agents that hold up: what we build in from the start

Tool contracts matter more than model choice, approval queues decay on a schedule, and the escalation rate everyone wants to drive down is worth protecting.

7 min readFilament
Architecture

The decision is the unit of work

Alerts, tickets, cases and documents are all proxies. Building around the decision instead changes what you measure and what you build.

5 min readLodestar
Platform

What changed recently

All release notes
2026.7

Platform release 2026.7

Evidence quorum becomes the default for consequential decisions, Frankfurt enters GA, and the trace store gains a query API.

2026.6

Platform release 2026.6

B200 capacity reaches GA, private Marketplace listings arrive, and evaluation gates can now block promotion on a per-case basis.

Developers

Documentation lives in the developer portal

API reference, SDKs, the CLI, the Terraform provider and working samples are versioned against your platform release. Access is included with every paid subscription.

Open the developer portalSee pricing