eight-week programmes that ship a tool the team uses on a tuesday. not a demo. not a pilot. the tool ships on day forty; the team owns it on day fifty-six.
the day-forty tool, not the day-zero demo
cadencefull-time pair one team, one ml-eng
term8 weeks fixed scope
shipped onday 40 in production
owned byyour team day 56
the work, named.
most ai engagements ship a slide deck and a slack channel. ours ship a tool on day forty — running in production, with a runbook, with the team using it. the next sixteen days are pair-time. then we close the laptop.
what we hold while we sit in.
the problem definition — what tuesday actually looks like, not what the demo looks like.
the data layer — feature store, evals, audit trail. nothing ships without an audit trail.
the model choice — open weights where it matters, hosted where it does not.
sovereign hosting where the data demands it. afrihost / hetzner / your tenancy.
the runbook — what to do when the model drifts, when it is wrong, when it goes dark.
the eval set — your own data, your own ground truth. shipped with the tool.
what we do not hold.
the strategy slide. you have one already. we read it; we do not write it.
the demo. we will not pre-record one for your board.
the moonshot. we ship what tuesday needs, this quarter.
a model we do not have evals for. period.
a budget we do not have your sign-off on.
§ posturethe test of an ai engagement is whether the audit team is using the tool on a tuesday — and would notice if you took it away.
how it runs.
eight weeks. four phases. one tool, in production, owned by your team. the runbook is the deliverable; the model is the by-product.
cadence
definew. 01
build the evalsw. 02–03
ship to one teamw. 04–06
hand-overw. 07–08
A.01 · define
one week. one document.
we sit with the team that will use it. we name the tuesday-task. we draft the eval set on paper. nothing is built yet.
A.02 · evals first
two weeks. nothing ships without them.
we build the eval set from your data, your ground truth. the model has to pass these before it sees a user.
A.03 · ship to one team
three weeks of build + pair.
one team. one model. one path through. the tool is in production on day forty. the team is on slack with us.
A.04 · hand-over
two weeks. then we leave.
the runbook is signed. on-call is named. the eval set is the team's. we close the laptop on day fifty-six.
this is for you if…
signals.
you have an internal team doing a repeated, structured task — audit review, claim triage, working-paper drafting, ledger reconciliation.
you have the data, in your own systems. not a demo dataset.
you can name the team that will use the tool. by name, not by department.
you would notice on day sixty-one if the tool went dark.
you want the tool owned by your team, not by a vendor api you cannot inspect.
not for you if…
you want a chatbot on your homepage. we do not do that.
you want a pilot that runs forever. we ship a tool; we leave.
you want to outsource the eval set. the eval set is yours; it stays yours.
you do not have the data. we can build a data layer first — see platform builds.
commercial shape.
fixed-scope eight-week sow. priced as a single block, billed at start and at day forty. halal-compatible commercial structure available.
8-week programmeone tuesday-task. one tool. one team. shipped in production on day forty.r 880kfixed scope
discovery-onlytwo weeks. eval set drafted, problem named, budget priced. no build.r 180kfixed scope
post-hand-over supportone half-day per fortnight for ninety days. they call us.r 32k/mooptional · 3 mo cap
example · a hospital group, claim-review ai.
a private hospital group in gauteng, sixty thousand monthly claims, three full-time reviewers running behind. eight-week programme. on day forty the tool was triaging the queue; on day fifty-six the reviewers were running it without us.
w. 01–02
define & evals.
shadowed two reviewers for a week. drafted four hundred labelled cases as the eval set. nothing built yet.
w. 03–06
ship to one team.
open-weights model in hospital tenancy. triages the queue; reviewers approve / overturn. eval f1 hit threshold on day thirty-six.
w. 07–08
hand over.
runbook signed. on-call rotation drafted. eval set is the hospital's. drift-alerts in their slack, not ours.