Skip to content
New guide: how to scope an AI project before you build

Our technology

How we build intelligent systems

We are a technology company. That means we care less about a particular tool and more about whether a system is grounded in your data, measurable in operation and changeable by your team.

The short version

The problem sets the technology. The system stays yours.

We start with the decision or process that needs to improve. Often a database query or an explicit rule is the right answer. We use AI where it earns its place, and we ground it in data you own and can check.

A model on its own is not a system. Useful intelligence needs a data model, permissions, an evaluation set, an interface and a way to recover when something fails. Those are software engineering problems, and we treat them that way from the first milestone.

The useful question is whether the result will still be changeable in two years. Most other decisions follow from that answer.

A good fit

When this approach works

  • The decision or process is specific to how your business operates.
  • The data exists, has an owner and can be used for the purpose.
  • Success can be measured with real examples.
  • Your team wants to own the code, the data and the infrastructure.
  • There are people on your side who can make product and data decisions.
  • You would rather build one system well than buy several tools that nearly fit.

A poor fit

When it is the wrong approach

  • A hosted tool already does the job well.
  • Nobody wants to own the system after launch.
  • There is no data, or no agreement on what good looks like.
  • The goal is a demo, not something operations will run.
  • The timeline assumes a managed platform’s operational load.

Ownership

Who runs the system

Decide who gets paged when something fails, and who can change the rules. Whatever the model or the cloud, someone has to own the system. That can be your team or a partner, but it should be written down before launch.

We build the operational baseline into the engagement: monitoring, alert thresholds, a deploy and rollback path, evaluation runs before each release, and runbooks written for someone who did not build the system. That is what makes ownership work in practice.

How an engagement runs

Questions

Before you commit

Do you only work with a particular model or vendor?

No. We choose the model and the infrastructure to fit the problem, the data constraints and the budget. The parts that matter most, such as retrieval, evaluation, permissions and the interfaces your team uses, are owned by you and can be changed independently of the model.

Where does our data go?

That depends on your constraints. Where data must stay in your environment, we design around local or private inference. Where an external API is acceptable, we document exactly what is sent and put access controls in place.

Can this run on our existing cloud?

Usually, yes. We work with Cloudflare and the major hyperscalers, and we plan around your existing systems of record rather than forcing a replacement. The setup is scoped against your constraints.

What if our team already uses another stack?

We build in mainstream languages such as TypeScript and Python. If your team works elsewhere, we document the boundaries and pair during handover so your engineers can change the system without us.

How do you know whether an AI system works?

We agree an evaluation set of real examples before launch and score the system against it. After launch we track the same measures, plus human overrides and cost. A system without an evaluation set is a demo, not a product.

Next step

Test the fit against your real problem.

Bring the problem, the data constraints and the three changes you know are coming. We will tell you what we think.

Prefer email? Write what you are working on and we will reply.

Send a note

By sending this, you agree that we may reply to your enquiry. See our privacy policy.