How we work

Small on purpose.
Governed by design.
One signature on the work.

Two founders who sign every engagement direct a governed AI workforce. No account managers, no blended-rate juniors, no handoff to a delivery team you never met. The gates between our work and you are written down — this page is that writing.

2
Founders who sign the work and answer for it end to end
9
Engineering seats, one written discipline standard each
4
Words we hand back instead of “should work” — the evidence states
1
Human decision between the work and your production

The shape

Smallness is a deliberate architecture, not a limitation to apologise for.

You buy a named founder’s judgement. What multiplies it is the workforce — nine engineering seats, one discipline each, working to a written standard from a shared, version-controlled corpus. The founders’ judgement is the scarce resource; the workforce applies it without diluting it. This is not a staffing model. It is an operating model, and it is running today. These nine seats are the ones a client engagement runs on; the wider estate — the AI board, the two lab teams and the Hermes delivery streams — is set out on the Factory Floor.

The line every piece of work crosses

Nothing reaches a client without a review gate, an integration gate and a human decision.

EAIC governed AI workforce A six-stage workflow from founder judgement through engineering, review and integration gates to a named human decision. 01 · FOUNDERSDecide theshapeSign the work.Set the question.Hold the outer gate. 02 · SENIOR SEATDecomposethe briefRead the corpus.Sequence the work.Commission seats.ONE BRIEF 03 · NINE SEATSBuild inparallelOne discipline each.Shared standards.Evidence on hand-back. 04 · REVIEW GATEDraft untilacceptedVerified.Built.Designed. Intended. 05 · INTEGRATEParts againstpartsNo part passes alone.Whole-system check.Rollback ready. 06 · HUMANAccept orescalateNamed person.Options attached.Decision recorded.
Nothing is deployed, published, purchased or made irreversible by an agent.
Founders

Decide the shape

Sign the engagement, settle open questions, hold the outer gate.

Senior engineer seat

Decomposes the brief

Reads it against the standards corpus and the decision log. The only seat that commissions work.

Nine seats

Build in parallel

One discipline each. Every hand-back carries an evidence state.

Review gate

Draft until accepted

Verified, Built, Designed or Intended on every claim. “Should work” is not a state.

Integration gate

Parts against parts

Nine parts that each pass alone is the normal state of a broken system, so integration is its own gate.

Human decision

Accept or escalate

Nothing is deployed, published or purchased by an agent. A named human answers for what leaves the building.

The disciplines, named

Every seat works to a written standard — not to a job title and a good mood.

Senior engineer

Reads the brief against the corpus and the decision log, decomposes it, sequences it, and commissions it. The only seat that commissions. Any seat may challenge its work — the objection has to be answered, not outlasted.

UI/UX engineer

Interface, interaction and accessibility — including the empty, loading and error states most builds leave out.

Middleware engineer

API design, interface contracts between services, performance under real load.

Data engineer

Pipelines, observability and data quality — so a wrong number is caught by the system, not by the client.

Database engineer

Schema, migrations and query performance, with the deprecation path for anything replaced.

AI/ML engineer

Models, prompts, evaluation and guardrails — evals agreed before the prompt code is written.

Security engineer

Hardening, insecure defaults, supply-chain risk — reviewed as its own workstream, not a checklist at the end.

QA engineer

Test-driven development, real browser testing, differential review of what changed against what was claimed.

DevOps engineer

CI/CD, automation and shipping — including the rollback, which is written down before the work starts.

Technical writer

Documentation, architecture decision records and the claims list — every factual assertion mapped to its source.

The review discipline

Everything an agent produces is a draft until accepted.

Every hand-back carries an evidence state on every claim. These four words are what we hand back instead of “should work” — and they are what you can hold us to, claim by claim.

Verified

Ran it, output shown

The command and its real output are in the hand-back. Not a description of the output — the output.

Built

Exists, committed, cited

It is on a branch at a named commit. You can go and look at it without asking anyone.

Designed

Decided, written down

The decision is made and recorded where you can read it. It is not built yet, and the hand-back says so.

Intended

Not started

Named as not started, on purpose, so nobody downstream builds on it.

The human boundary

Agents propose, build and verify. Humans decide anything a reasonable person could disagree about. Positioning, spend and external commitments are a human decision — so is anything irreversible, legal, regulatory or safety-critical. Agents do not merge to default branches, deploy to production, delete data, or commit anything that speaks outside the box. Disagreement is escalated at its strongest, not averaged away; escalation reaches a human with options and a recommendation attached, never a bare question.

When a seat is wrong

Work found wrong is rejected against the written definition of done. A seat that is blocked names its blocker and the owner who can clear it. Any seat may challenge any other seat’s work — including the senior engineer’s. The objection has to be answered, not ignored.

What a client actually sees

The same artefacts the operation runs on.

Governance is the answer to “how do I know an AI built this properly?” A buyer who asks gets a documented answer — not a testimonial, not a logo, and not a guarantee without a methodology behind it.

The decomposition

Every workstream, its dependencies, and its acceptance criteria.

The review states

Verified, Built, Designed or Intended on every claim in every hand-back.

The claims list

Every factual assertion mapped to its source, for line-by-line approval.

The decision log

What was decided, by whom, when, and why.

The rollback plan

Written down before the work starts, not after something breaks.

The escalation record

What was escalated, with the options and recommendation that went with it.

Why this matters to you

Speed without evidence is just risk with a schedule.

The workforce makes us fast. The review discipline makes the speed safe to buy. You get weeks-not-quarters delivery and an audit trail that still answers the question a year later — and the two are the same system, not a trade-off.

See what you can commission

What this page does not claim

No counts of systems shipped, no speed or cost multipliers, no client references. Where we cannot evidence a number, we do not print one. Everything here is present-tense and verifiable now; a track record is not claimable without evidence, so this page does not claim one.

Want this run on your problem?

Tell us what you are trying to build and we will tell you how we would decompose it — before you commit to anything.

Price agreed before we start · Founder-led · NDA available