Start a project

How the work runs

The same four phases whatever the project is. The point of running it this way is to force the decisions into the open early, because the ones nobody makes are the ones an AI tool will quietly make for you, in code, where they are expensive to find.

Four phases

  1. Clarify

    Week 1

    We find the decisions nobody has made yet, the ones an AI tool would otherwise make for you, silently, in code. You get a written scope and a fixed price before anything is built.

  2. Plan

    Weeks 2-3

    Architecture, data flow, agent boundaries, and what counts as working. All of it agreed while it is still cheap to argue about, and written down so nobody relitigates it in week six.

  3. Build and verify

    Weeks 3-7

    Small changes, each one run against the evaluation set. You get a live environment in the first week and watch it fill in, so nothing arrives as a surprise at handover.

  4. Learn and adjust

    Week 8 onward

    Real use produces cases the plan missed. Those become new tests, the system improves against them, and the loop keeps running after handover, with your team or with ours.

What every change travels through

Phases describe the engagement. This describes a single change inside it, from the moment someone edits a prompt to the moment it is serving everybody.

  1. 01 change

    Something moves

    A prompt, a retrieval config, a model version, a line of code. In an AI system all four ship the same way, which is the first thing most teams have to fix.

    Nothing reaches this pipeline without a review and a reason.

  2. 02 suite

    The evaluation set runs

    Real queries from your staff, with expected behaviour recorded. Not a public benchmark, which tells you how a model does on somebody else's problem.

    A score below the agreed floor stops here and never reaches staging.

  3. 03 shadow

    It runs beside the live system

    The new version answers the same traffic as the old one without anybody seeing it. Disagreements between the two are the interesting output.

    Divergence past the threshold holds the release for a human to read.

  4. 04 flag

    A few per cent of real traffic

    Behind a feature flag, on a cohort you choose. Latency, cost per call and the failure rate are watched against the version still serving everyone else.

    Any of the three moving the wrong way flips the flag back automatically.

  5. 05 rollout

    Widened in steps

    Ten per cent, fifty, all of it, with a soak at each step long enough for a slow problem to surface. Fast rollouts hide exactly the faults worth catching.

    The previous version stays warm and addressable for the whole rollout.

  6. 06 watch

    The set grows

    Anything that went wrong in production becomes a case in the evaluation set, so the same fault cannot ship twice. This is why the loop is a loop.

    A rollback is a rehearsed command, not an incident-time decision.

What you can hold us to

Evaluation before opinion
Every AI system we ship has a test set drawn from your real queries and a score you can track over time. When we claim something got better, the number comes with it.
You own all of it
Source code, prompts, evaluation sets, infrastructure and accounts transfer to you on final payment. Nothing stays licensed from us.
No model lock-in
The model sits behind an interface. Swapping it is a config change and an eval run, so you are never stuck with a provider because leaving would cost too much.
We will talk you out of it
If your problem is better solved by a form, a query or a rule, we will say so before you have paid for a model to do it worse.
Fixed scope, fixed price
You approve a written scope and a number before work starts. Changes get quoted, not absorbed silently and billed later.
Built to be handed over
Every project ships with documentation and a training session. If you replace us next year, the next team can read what we left.

Start with the problem

The useful first message is a paragraph about what is not working, not a specification. We come back within two working days.

Start a project