Docs

How we build, written down.

The working guides we hand to client teams: models, infrastructure, integrations, evals and governance. Free to read, no gate.

guide published

areas covered

Monthly

updated as we learn

Free

no signup to read

Start here

Ship your first AI workflow

The full six-step sequence with worked examples, process maps, eval output and integration code, plus the readiness questions to clear before step one.

Read the guide

Working with language models

Practical guidance for teams putting Claude and other models into real workflows.

Choosing a model for a taskReasoning depth, latency and cost are a trade, not a ranking. How we decide per step.8 min Writing prompts that survive handoverPrompts as versioned artefacts with tests, not strings buried in code.7 min Retrieval that respects permissionsChunking, citation and row-level access so answers can be checked and defended.9 min When to fine-tune, and when not toAlmost always not. The three cases where it earns its keep.7 min

AI infrastructure

How we deploy, size and operate systems in your cloud, on-premise or air-gapped.

Reference architectureThe standard shape: gateway, orchestration, tool layer, audit store.10 min Sizing from 0 to 1M+ usersWhat actually changes at each order of magnitude, and what does not.8 min Self-hosting open weightsHardware, quantisation and the honest performance gap on your own metal.9 min Cost controls that holdPer-run budgets, caching and the three places spend quietly escapes.7 min

Integrations

Getting output into the systems your business actually runs on, reversibly.

Idempotent write-backWriting into a system of record without creating duplicates or orphans.8 min Legacy adaptersSOAP, SFTP drops, fixed-width exports and the patterns that tame them.7 min Exposing systems as safe toolsGoverned tool definitions so the next agent plugs in instead of being rebuilt.9 min Queues, retries and backfillsThe unglamorous plumbing that decides whether a pilot survives month two.8 min

Evaluation and quality

Proving it works, before rollout and on every change after.

Building an eval set from your historyHow to assemble a scored set your own team trusts and can rerun.8 min Choosing the metric that mattersAccuracy is rarely the number the business cares about.6 min Regression gates in CIBlocking a deploy on quality, not just on tests passing.7 min

Security and governance

What your security review will ask, and how the system answers it.

Audit trails for AI decisionsWhat to log so an auditor can reconstruct any decision the system made.8 min Designing approval gatesWhere a human must sign off, and how to place gates without killing the gain.8 min Data residency and sovereigntyKeeping data, inference and processing inside your own boundary.9 min Rollback and kill switchesEvery automated action undoable in one step, owned by your team.7 min

Operations playbooks

The patterns we reuse, written up so your team can run them without us.

First pass, human approveThe fastest safe way into a regulated or high-consequence process.7 min Exception routingAutomate the routine path, escalate the rest with reasoning attached.8 min Running an AI system on callRunbooks, alerts and what to do at 3am when quality drops.8 min

Want these applied to your operation?

The guides are the method. The Audit is the same method pointed at your systems, with a build plan at the end of it.

Schedule call