Four phases, no pilots, designed to operate for a long time.
Our method is deliberately unfashionable. We do not run pilots, we do not hand off to juniors, and we expect to still be involved with the systems we build a year after they launch. The phases below are how every engagement is run.
- PHASE I · 2–4 weeks
Diagnose
We begin every engagement by studying the problem without selling against it. Structured interviews with the operators who will live with the system; a read of the data and infrastructure as it actually is, not as it is documented; and a written thesis about what we have learned.
- 01
Primary sources first. Every engagement begins in conversation with the people doing the work.
- 02
We read the code and the data before we write the plan.
- 03
The deliverable is a thesis, not a proposal. If the thesis is that you should not build the thing, we will say so.
- 01
- PHASE II · 2–6 weeks
Design
Architecture, evaluation strategy, rollout plan, operating model, and a costed engagement. Every design is reviewed by a principal outside the engagement before it is presented.
- 01
Evaluation is designed alongside the system, not after it.
- 02
Cost, latency, and failure modes are first-class inputs to the design.
- 03
Designs are reviewed by a second principal. Nothing ships without a second look.
- 01
- PHASE III · 6–16 weeks typical
Deploy
The first deployment is scoped narrowly but built to production standard from day one. Observability, on-call, and evaluation harnesses are in place before the first real user touches the system.
- 01
No pilots. The first system we ship is a production system, or we have not shipped.
- 02
Ship the runbook with the system. Operators should not discover how to run the system in anger.
- 03
Ship behind a feature flag. We iterate under production load, not under demo conditions.
- 01
- PHASE IV · Quarterly cadence
Operate
We stay. Most engagements continue into a quarterly operating cadence: model and system upgrades, observability review, new evaluation sets against new distributions, and a standing seat for the judgment calls that arise.
- 01
A system is not finished when it launches. Most of the value is earned in year two.
- 02
Review cadence is calendared, not discretionary.
- 03
We design for handover to your team — and we remain on retainer when you want us to.
- 01