Ship real software with AI. Own
We help engineering teams ship with AI under engineering control, at a cost you can measure and defend. And we prove it with receipts.
What efficient AI delivery is made of
60%
software engineering best practices
40%
AI tooling fluency
0
claims without a receipt
[ 01 — The problem ]
One problem, three facets
You lead a team of 5–100 developers shipping real software under AI pressure. The board wants visible productivity. Finance wants cost accountability. Engineers don’t trust AI output enough to own it. And the AI tooling bill arrives with no attribution.
[ mi review · sample ]
$ mi review --task PAY-142
diff +214 −38 · 6 files
tests 11 passed · 2 added
evidence trace + screenshots attached
review approved · risk noted
PR owned by a human ✓ [ legroom report · sample ]
$ legroom report --week
spend (API-equivalent) $184.20 exact
cache reads / input 71.4% exact
saved by caching $312.55 estimated
attribution per task ✓ [ mi install · sample ]
$ mi install --team payments
week 1 workshop · 8 devs
week 2 4 tasks through the loop
week 3 baseline + policy pack
week 4 playbook handed off ✓ Sample receipts. The real ones come from your runs — every number labeled exact or estimated.
“Efficient AI delivery is roughly 60% software engineering best practices and 40% AI tooling fluency.”
— the base model — everything on this page follows from it
[ 02 — The method ]
AI-assisted delivery is continuous delivery, extended
AI does not replace your delivery pipeline; it accelerates the inside of it. The same gates, a radically faster loop between them. Trust lives at the gates; speed lives between them.
The delivery loop — ship/compound feeds back into prepare
Three loops, three timescales, strictly nested
| CD gate chain | org | days–weeks | where a change may stop |
| delivery loop | task | hours–days | one task traversing the gates |
| AI inner loop | change | minutes | route → change → check → evidence |
[ 03 — What we do ]
Pick your rung
Every engagement ends with the capability installed in your team — not a dependency on us.
AI delivery workshop → install
Module one is a hands-on workshop: why delivery breaks under unmanaged AI, the loop, task selection, context packs, verification-first work, reviewable diffs. Weeks 2–4: your developers ship real tasks through the loop while your AI spend gets a meter and a baseline. The playbook stays with your team.
Book a diagnostic call →Forward-deployed AI engineering
An engineer embedded in your organization, shipping through the loop on your real backlog and installing the practice as we go. For enterprises that want the capability built inside, proven on production work — not slideware.
Book a diagnostic call →Legroom
The local-first meter and guard for agentic coding: wire-truth observability of what your AI tooling actually does and costs, guardrails against invisible waste, and evidence-verified routing. The instrument behind every number we quote.
legroom.dev ↗Berek
AI does the work, you decide — the company’s AI collaborator on your own infrastructure. Agents and your team share sandboxes, integrations, and workflows; every task reaches a human decision with evidence, and the accumulated knowledge stays a company asset. Pairs with Legroom for measured, replaceable model routes. Early product, looking for 1–3 design partners.
berek.dev ↗[ 04 — Principles ]
How we work
- engineering AI engineering is software engineering.
- economics Cost is the door; control is the product.
- receipts No claim without a receipt.
- gates Trust lives at the gates; speed lives between them.
- ownership Human ownership stays central: reviewable diffs, human-owned PRs.
- focus Cut anything that does not compound.
What did we ship? · What did we prove? · What did it earn? · What did we teach?