Process

Five steps. No pilot purgatory.

You don't buy a big custom project on a hunch. You move through five steps, see working software every Friday, and decide at every one.

01 · AUDIT Weeks 1–2 02 · DESIGN Weeks 3–4 03 · BUILD Weeks 5–8 04 · PROVE Weeks 8–10 05 · RUN Ongoing fix & re-run · evals decide

The pipeline — work only moves right when the evals say so.

01

Weeks 1–2 · Fixed price

Audit

Two weeks inside your operation. Owners rarely bring us a problem — they bring a solution ("build us a chatbot"). The audit digs under the stated request: we interview the people who actually do the work, map how it really flows (including the parts nobody documents), and score every process by the value at stake and how much AI can genuinely do. If the audit surfaces nothing worth automating, you don't pay for it.

The method

Reframe the stated requestMap the undocumented workflowScore value × AI fitOne list: now, later, leave alone

You walk away with

A complete process inventory

A ranked roadmap with the ROI math shown

A fixed quote for the first build — no estimates

02

Weeks 3–4

Design

Each workflow specified end to end before a line is built — with the outcome defined in your numbers, not our activities, and a clear "done" test on everything in scope. We draw the code-versus-model line so the right component makes each decision, and governance is designed in from day one: what the system may do, touch and act on, with escalation paths and named owners.

The method

Outcome in your numbersA "done" test on every scope lineCode vs. model decision boundaryGovernance as design requirements

You walk away with

End-to-end workflow specifications

A data-flow map your IT team can audit

Escalation rules — every exception has a named human

03

Weeks 5–8

Build

We build and test against your real data, in your stack, on infrastructure you control. We engineer the context the model needs, draw the code-versus-model line so the right component makes each decision, and map the situations real users will throw at the system — before they happen in production.

The method

Context engineered, degradation diagnosedCode vs. model per decisionReal-world scenarios mapped up frontReliability practices, not vibes

You walk away with

Production systems, tested on your real data

A working demo every Friday

A plain-English runbook and every credential

04

Weeks 8–10

Prove

The step most projects skip — and the reason they quietly die six weeks after the demo. We define what "working" means in measurable terms and prove it: deterministic checks, model-graded review and human review, plus the economics — what the system costs to run against what it returns, in a business case a finance-minded decision maker signs off on.

The method

"Working" defined measurablyEvals: deterministic + model-graded + humanRun-cost vs. value returned, modelledFix, accept, or call it done — by the numbers

You walk away with

An eval suite that keeps scoring the system

A cost-and-impact report in your numbers

The go / no-go call before anything scales

05

Ongoing · Optional

Run

A system nobody uses is a system that failed — so delivery is designed around adoption, not bolted on at the end. Your team is trained, the feedback loop catches drift before your customers do, and usage is tracked honestly. Then we stay only as long as you want us.

The method

Adoption designed, not assumedDrift caught before customers noticeTraining, docs and enablement includedUsage tracked, honestly

You walk away with

Monitoring dashboards and alerting

Team training and handover

Quarterly reviews — or a clean goodbye

You decide at every step

Each step ends with a deliverable you could act on with or without us. No long contracts, no sunk-cost traps — you approve the next step knowing the math.

You own everything

The systems, the credentials, the documentation, the trained team. We build on tools you own, with no proprietary lock-in and nothing rented back to you.

Working software every Friday

From week five you see the real system running on your real data, every week. If a week produces nothing you can click, we've failed that week.

Step one takes two weeks.
The call takes thirty minutes.