Loading page
Five stages from a paid audit to an agent carrying live traffic, each one closing on a document, a demo, a passing eval or a dashboard. Nothing carries over on trust alone, and you can stop at the end of any of them.
Stages one and two are fixed price. Stop after either and you keep everything we produced.
We build and operate AI systems that reach production.
The same sequence on every engagement, whether it runs nine weeks or nine months.
A paid two-week audit of your processes, data and systems. We come back with a ranked list of what AI should touch — and what it should not.
Ends withAudit report and ranked backlog
Model selection, retrieval design, guardrails and the integration surface. You get a written architecture with a cost model attached to it.
Ends withArchitecture doc with cost model
Two-week sprints with a working agent at the end of each one. Your team reviews real output, not slideware, from the third week onward.
Ends withA working agent every sprint
A task-specific eval suite gates every release. Nothing reaches production until it beats the baseline on your data, not a public benchmark.
Ends withEval suite and release gate
Tracing, drift alerts, cost dashboards and an on-call team who already know your stack. We stay after go-live — that is the whole point.
Ends withLive dashboards and on-call rota
Claude writes the code and composes the interfaces. It does that against a locked architecture and inside a design system that already exists — which is the difference between generation that compounds and generation you throw away.
Stage two produces a written architecture — schemas, tool boundaries, authority limits. That document is the context every generation runs against, so the code answers to something reviewable rather than to whatever was typed into a chat that morning.
The type scale, spacing and colour tokens are defined before any screen is. Interfaces are composed inside that system, which is why a generated screen looks like the rest of the product instead of like a demo bolted to the side of it.
Every diff is read by the engineer who will be on call for it. Nothing merges because it looked plausible; the person approving it is the person who gets woken up when it is wrong.
What this changes is the cost of getting to a reviewable draft. The eval suite, the trace requirement and the merge standard are exactly what they were before — the bar did not move, the run-up to it got shorter.
A real tool definition, not pseudocode. The lines that matter are the authority limit, the required trace and the two escalation paths — all four come from the architecture document, and none is the model’s to choose.
Generation is fast enough to outrun review, so the constraints live in the repository rather than in anyone's head. One committed file, read by every assistant that touches this codebase, checked again by CI on the way in.
What it refuses
A rule that only lives in a review comment is a rule that holds until the week everyone is busy. These fail the build, which is the only enforcement that survives a deadline.
ci · rules check
1 blocked · build failed · nothing merged
A typical engagement to production. Evaluation starts the same week building does, and neither it nor operations has a finish line.
Week
Indicative, not contractual
Fixed windowRuns on past go-live
Three commercial shapes, one per phase. None of them is an hourly rate.
Fixed price
Two weeks, agreed before it starts. You keep the audit and the ranked backlog whether or not you continue with us.
Fixed scope per phase
Each phase priced against the architecture that discovery produced, so the number is attached to something written rather than to an estimate.
Monthly retainer
Scaled to traffic rather than to headcount. Tracing, drift alerts, cost dashboards and a place on the on-call rota.
Each one has cost us an engagement we could have won. That is how you know they are real.
Discovery ends with a list of what to automate and a list of what to leave alone. The second list is the one that earns the fee, and we deliver it in writing whether or not you like it.
Nothing reaches production because it impressed a steering committee. If we cannot measure it against the work it replaces, on your data, it does not go live.
There is no version of this engagement where we build a system and walk away from the pager at go-live. Handover without operations is how these projects quietly die.
Hourly billing pays us to be slow. Every phase carries a number agreed before it starts, and no change enters scope without a written estimate first.
Bring one process you think an agent could run. We'll tell you straight whether it's worth building — and what it would cost if it is.