Discovery ends with a proposal. What follows is the longer part, and most of what decides whether it goes well is unglamorous: who is in the room, what happens each week, what each side owes the other, and what you are holding at the end.
The four ways we engage
We are not a body shop and we are not only an outsourcer. Which mode fits depends on what you have and what you are trying to build.
We deliver it. You bring the requirement, we assemble the team the problem requires and deliver the result. Suited to organisations without an internal AI team, or with one that is fully committed elsewhere.
We embed in your team. Our people work inside your existing team, in your repositories, your review process, and your rituals. Best when you have good engineers who have not built this kind of system before, and you want the capability to stay with them.
We lead your team. We take technical leadership of an existing team that has the people but not the direction: setting the approach, establishing the evaluation and engineering practices, and unblocking work that has stalled.
We build your team, then leave. We start by delivering, and along the way we help you hire your own people, interview candidates with you, onboard them onto the system we built together, and reduce our involvement in stages as they take it over. This is the most satisfying version of the job and the one that most obviously ends with us out of a contract, which is the point.
Modes change during an engagement. Delivering often turns into embedding, and embedding often turns into building your team. We would rather change the arrangement than pretend the original one still fits.
The service pages describe what these look like in practice: AI transformation and strategy, custom AI development, team training, and AI engineering workflows.
Whichever mode we are in, the people are the ones named in the proposal. That is covered on the previous page.
The weekly rhythm
Process should be as light as the project can get away with, and no lighter. Our default:
A short daily sync. Ten to fifteen minutes, on a call or in writing. What each person is doing, and what is blocking them. Blockers get resolved immediately after, not in the meeting.
One prioritised queue. A single ordered list that everybody, on both sides, can look at to know what is next. Not one list per person, and not a set of documents that disagree with each other. It is the canonical record of the plan, which means decisions made in a call get written into it or they did not happen.
A weekly review. Once a week we look at what we learned, what shipped, what changed as a result, and what is coming. Everyone says how they feel about the week, including the uncomfortable parts, and every risk anyone can see gets named out loud before we work out together what to do about them. We write the notes down and share them.
In a research phase the weekly review is about findings rather than features: which hypotheses survived, what the numbers say, what we now believe that we did not believe last week. The obligation is the same. Something real, every week.
Decisions get written down. Choices with consequences, such as an architecture, a model family, an evaluation metric, or a tradeoff we accepted, get a short written record: what we chose, what we considered, and why. It takes ten minutes and it saves an argument in month four, when the person who remembers the reasoning is on another project.
Process adapts. If a ceremony stops earning its time, we drop it. If a project needs more structure, we add it. We would rather change the process than keep it out of politeness.
What we need from you
Our half of the work depends on yours.
A decision maker. One person who can settle a question about priority or scope without a committee. Not necessarily senior, but genuinely empowered.
Access, early. Data, systems, and the people who understand them. This is the single most common cause of a slow start, and it usually needs somebody to push it through legal and IT before the engagement begins.
Subject matter expert time. A few hours a week from someone who actually understands the domain. We can learn your business, but not as fast as you can explain it, and the difference shows up directly in the quality of the result.
Regular acceptance. Look at what we produce while it is being produced, not at the end. Tell us when it does not match what you expected, early enough that changing it is cheap.
Honest constraints. Real budget, real deadlines, real regulatory limits, real politics. We plan better with the awkward facts than without them.
What you own
Everything we build for you is yours: code, model weights, datasets, evaluation suites, documentation, and infrastructure definitions. Yours as we go, in your repositories and your accounts, rather than handed over in a bundle at the end.
We do not build on a proprietary platform of our own, and we do not keep a critical piece in an account only we can access. A handover package can be requested at any point in an engagement rather than only at its conclusion, and it is never withheld as leverage. The one condition is the ordinary one, that invoices are settled.
We sign mutual NDAs. We talk to a lot of companies, so ideas that sound unique often are not, and we will say if we see a conflict with existing work rather than sign and hope.
How an engagement is structured commercially is a conversation rather than a page. Commitment, cadence, and cost depend on the problem, and we would rather work them out with you than publish a shape that fits nobody.
After handover
AI systems decay in ways ordinary software does not. Data drifts away from what a model was trained on. Providers deprecate the model version you built against. Prices and rate limits move. The evaluation set that was representative last year quietly stops being representative.
So handover is designed for your independence. You get runbooks, the procedure for retraining and re-evaluating, alarms for the things that drift, and a documented cost model, so that a surprise on the invoice is not the first sign of a problem.
From there it is your choice. Take it in house entirely. Keep us for a periodic evaluation refresh and model upgrades. Bring us back for the next problem. All of those are normal, and none of them are enforced by making the system hard to run without us.
The next two pages describe the craft underneath all of this: how we run research so its results can be trusted, and how those results become a system you can operate.
Next: Research and Evaluation
Not sure which mode fits you?
Book a free 30-minute call. Tell us what you already have in-house and we will tell you which of the four we would suggest — including when the answer is none of them.
No pitch deck, no obligation. If AI is the wrong answer for your problem, we'll tell you that too.