01Week 1
Shape
A paid week at a fixed price. We work out what the thing has to do, run the hard cases against your real data, agree an approach, and you get a fixed quote for the build.
AI engineering studio
AI that still works at question eleven.
Most AI features pass the demo and fail in production. We build the half that has to hold.
Start a projectsystemsInterfaces
around them
Scroll to begin ↓︎
The demo works.
Ten questions in, everyone is convinced. It answers cleanly, the room nods, and the project gets its budget.
Question eleven is where it breaks.
A real user asks something the demo never covered, and the model answers confidently and wrongly. Worse: nobody can say whether last week’s change made things better or only different.
We build the half that has to hold.
An evaluation set before the first prompt. Answers grounded in your own material. A person on anything the model is unsure about, tracing on every call, and a way back when something breaks at two in the morning.
Saılbuıld
AI engineering studio
Chapter 1
AI solutions
Grounded answers
Every claim links back to its source.
Retrieval across your own documents, records and tickets, with the passage it came from one click away. Retrieval is scored on its own, so a wrong answer can be traced to the right cause.
Agents that hand over
Anything ambiguous goes to a person.
They call the tools your team already runs, and escalate rather than guess when the situation is unclear. Every step is traced, so you can see what it did and why it did it.
Measured before it ships
An evaluation set before the first prompt.
Cases built from your real data, scored on every change and run in CI. When a prompt edit or a model version makes things worse, you find out before your users do.
Chapter 2
Websites
One voice
Sites, product interfaces and dashboards.
A design system underneath them all, so every surface your customers meet speaks the same way.
Yours to edit
Change a headline without filing a ticket.
A CMS your own team can use, so the site keeps up with the business without waiting on a developer. Fast on a mid-range phone and readable by a screen reader.
The states nobody designs
Waiting, citations, corrections, “I don’t know”.
AI interfaces live or die on the unhappy paths. They get designed here, alongside the system, rather than patched in the week before launch.
Chapter 3 ( How a project runs · six to ten weeks )
Four stages,
Priced before you commit,
shown working before you pay for more.
one price.
02Weeks 2–4
Design
Drawn in the browser rather than in slides. On AI work that means designing the states nobody enjoys: waiting, citations, corrections, and the model admitting it does not know.
03Weeks 4–9
Build
Weekly increments on a staging site you can open whenever you like. You watch the thing working before you pay for the next week of it.
0430 days
Launch & hold
It goes live, runs under real traffic, and we fix what real traffic finds. Thirty days, at no extra cost.
A paid week to shape it, then a fixed quote that cannot move. Weekly increments on staging, and thirty days of support after launch.
Book a callChapter 4 ( What it costs )
Three parts, and you know the price before the second one starts.
- 01ShapeOne week, fixed price
A paid week that ends with an evaluation set built from your own data, the approach written down, and a fixed quote for the build. If the honest answer is that we are not the right studio for it, you get that in writing and owe nothing further.
- 02BuildSix to ten weeks, fixed price
Quoted at the end of the Shape week and fixed from then on, so the number cannot move under you. You watch it working on staging every week before you pay for the next one.
- 03CareMonthly, optional
For after the thirty included days, if you want us to stay on it. Cancel with thirty days’ notice: everything is already in your name, so nothing is locked to us.
A paid week that ends with an evaluation set built from your own data, the approach written down, and a fixed quote for the build. If the honest answer is that we are not the right studio for it, you get that in writing and owe nothing further.
- The task written down in terms that can be tested
- An evaluation set built from your real cases
- Architecture and provider choice, with running costs estimated against your volume
- A fixed quote for the build, held for sixty days
Quoted at the end of the Shape week and fixed from then on, so the number cannot move under you. You watch it working on staging every week before you pay for the next one.
- Designed in the browser rather than in slides
- The build, the integrations and the interface people use
- Tracing on every call, hard spending limits, and a rollback that works
- Thirty days of support after launch, included
For after the thirty included days, if you want us to stay on it. Cancel with thirty days’ notice: everything is already in your name, so nothing is locked to us.
- Uptime and cost monitoring, with alerts that reach a person
- Evals rerun as your cases change
- Model and dependency updates
- A day a month of changes
Not in the price: model usage, billed to your own provider account at cost; third-party licences; and hosting after launch, which runs in your accounts. We quote in pounds or dollars, and say which before you commit to anything.
Chapter 5 ( Why it holds )
Built to be checked, not just demoed.
- 01
Evaluated, not argued
0%of 212 cases passEvery change is scored against your real cases, so progress is something you can see rather than something we claim.
- 02
It knows when to stop
0below the line →︎ a personConfidence thresholds and a plain “I don’t know”. Anything under the line goes to a person instead of being guessed at.
- 03
Costs you can see
0%of this month’s ceilingTracing on every model call, spending limits that cannot quietly run away, and a way back to the last version that worked.
- 04
Inside a latency budget
0sp95, question to first tokenSpeed is agreed as a budget at the start and held to, not measured once at the end. The same goes for the interface it arrives in.
Below the waterline
What your security review looks at is under the surface.
Everything above the water is the part you see working. What holds it up is next, answered here rather than eight weeks into procurement.
Chapter 6 ( Your data )
Does our data train a model?
Never.
Training opt-out is configured by default on every provider we call, and the same commitment flows down to the model vendor. It sits in the contract, not only in a settings page.
Where does it run?
Where you say.
The system runs in a UK, EU or US region on Azure or AWS, whichever your policy requires. You are told which one before anything is deployed.
Does failover cross a border?
It does not.
Failover runs across availability zones within the region you picked, so a hardware failure never moves your data over a border.
Is there a data processing agreement?
As standard.
Offered with every engagement rather than produced on request, with the sub-processors named inside it.
Who can touch our data?
It is written down.
Every service that can touch your data, from the model provider to hosting and monitoring, is listed and versioned. You hear about a change before it happens, not after.
What did the system do in March?
There is an answer.
Every model call is traced and retained on your terms, so when someone asks what the system did in March, there is an answer rather than a guess.
Answered here rather than eight weeks in. If your reviewers need this on letterhead instead of a web page, ask and it comes back the same day.
Chapter 7 ( Who it is for )
For every team that has to ship.
01
For founders.
A first AI feature that has to survive the demo and then the week after it.
02
For product teams.
AI inside a product that already exists, including the states nobody planned for.
03
When it already broke.
A feature that shipped and is now quietly answering wrongly. We measure how often, then fix it.
04
For operations.
Extraction and review queues that work through themselves and only ask when they are unsure.
Chapter 8 ( Before you ask )
The questions we always get.
Or ask us one of these
- Is our problem even an AI problem?
- Ours is already live and getting things wrong. Can you measure it?
- Can it run on our own data without leaving our systems?
- What would this cost at our volume?
Answered in one working day, usually less
Will it make things up?
Any model can. We ground answers in your own material and show the source, check confidence before anything reaches a user, and send whatever falls below the line to a person.
What if we already shipped something and it is going wrong?
That is the most common reason people call. The Shape week starts by measuring it against an evaluation set built from your real traffic, so you get a number for how often it is wrong before anyone argues about a fix.
Do you build websites on their own?
Yes. A good share of our work is only the site or the interface, with no model anywhere near it.
Who owns the code?
You do, from the first commit. Repository, hosting, provider accounts and domains all sit in your name.
What will it cost to run?
We estimate model spend during the Shape week against your real volume. Everything ships with hard spending limits and tracing on every call, so nothing runs away quietly.
How long does it take?
Most projects run six to ten weeks from the Shape week to launch. A site on its own is usually quicker.
What do you need from us?
Someone who can answer questions about the work, access to the content or data it runs on, and one person who can sign things off. About an hour a week once the build is moving.
Is our data used to train anyone’s model?
No. Training opt-out is on by default for every provider we call and the commitment flows down to the model vendor in the contract. You pick the region it runs in, and the sub-processors are listed rather than buried.
Can we start small?
That is what the Shape week is for. One week, one fixed price, and it ends with a quote you are free to walk away from. Nobody has to bet a large budget on a first meeting.
Why is there no price list?
Because a band wide enough to cover every job tells you nothing, and we would rather quote your actual scope than hedge. You get a real number on the first call, and a fixed one at the end of the Shape week.
Start something
Book half an hour.
Thirty minutes with the engineers who would build it. You leave knowing whether this is a job worth doing and roughly what it costs. If it is not something we should take on, we will say so on the call.