How we get AI into production.
Five things we do on every engagement, and the client project where each one shows. They are the reason we can work in social care, custody banking and fine instruments in the same quarter.
Start with the decision being made
What the technology can do is rarely the constraint. The constraint is what someone can act on, under pressure, with the information in front of them. We settle that before anyone chooses an architecture.
Propolis
Three audiences with conflicting commercial interests in one community. The structure was settled before a single screen.
Agree what a wrong answer costs
We work out the cost of being wrong before we design what right looks like. Where that cost is high — a missed decline in a resident, an incorrect provenance record — it sets the whole architecture.
Dot
The system's limits around vulnerable residents were agreed before any features were designed.
Put the audit trail on screen
Enterprise buyers have to defend an answer to a board. If they cannot trace it back to source in the interface, they will not use it, however good the answer is.
EVOLVE
Every answer shows its datasets, its assumptions, and which figures are observed rather than modelled.
Write down the decisions we push back on
Where we argue for something contentious — removing a dashboard, keeping AI away from a child's writing — we record why. That is what keeps the decision in place after we leave.
Stay after launch
Retrieval quality drifts, data sources move and users change how they ask. Launch is the start of the measurement period, not the end of the job.
Four phases, and you can stop after any of them.
Feasibility
A short, paid piece of work: discovery with real users, a technical read, the regulatory and duty-of-care position, and costed options. You get a decision you can take to a board, including where that decision is not to build.
Definition
Information architecture, failure modes, the data the system is allowed to touch, and the decisions a human must keep. This is where most of the risk is removed.
Design and build
Design and engineering run as one team. Interfaces, design system, integration and release — with accessibility designed in rather than audited afterwards.
Run
Measurement against the baseline we agreed at the start, then a standing team keeping it honest as the data and the users move.
What we build with, and why it matters to you.
We are not tied to one model or vendor. We pick the combination that gets the result at a cost you can carry, and we tell you what it will cost to run before we build it.
Retrieval over your own records
Answers come from your documents and data, with the source shown next to them. This is what made a 130-year archive at Beare's searchable in seconds.
Prompt and instruction design
The wording that sits behind each answer is written, tested and version-controlled like any other part of the product, so results stay consistent as usage grows.
Fine-tuning where it earns its place
We tune a model only when retrieval and instruction design cannot get there. Cheaper to run, easier to explain, and less to unpick later.
Conversational interfaces and assistants
Assistants that hold context across a task, hand over to a person at the point it matters, and log what was asked and answered.
Text, image and voice together
Where a job involves photographs, documents and speech, we handle all three in one workflow rather than three disconnected tools.
Forecasting and analysis
Models that turn existing data into a figure someone can act on — the work that took EVOLVE's analysis from six weeks to same day.
Process automation
The repetitive steps between systems handled automatically, with a person kept on the decisions that carry risk.
Bias, safety and audit
Limits agreed before build, checks in the release process, and an audit trail on screen so a decision can be defended to a board or a regulator.
Tell us what you are trying to fix.
Thirty minutes with the people who would do the work. You will leave with a view on whether it is worth building and roughly what it costs.