Loading
The machinery
A factory you can't inspect is just a black box with better marketing.
Every firm in this category now says the same sentence about AI. The difference is whether there's a line underneath it or a subscription to somebody else's chatbot.
Ours is six systems, built over two years, that we run every day on our own businesses before we point them at yours. Here's what each one does and why it's there.
01Build
Fat Controller
The orchestrator
Dispatches work to a fleet of AI agents running in parallel, each in an isolated git worktree. Six providers behind one queue. Every task captures its tool calls, token spend and cost, runs post-agent quality gates, and opens a pull request for review.
This is what makes the line a line rather than one person typing faster. Work is queued, allocated, sandboxed and measured — the same properties you'd want from any production system.
02Intake · Spec · Build
EOS
The memory
A context server holding every project's state, architecture decisions, conventions, runbooks and session history, searchable across the whole portfolio. Agents read from it at the start of a task and write back at the end.
The failure mode of AI-assisted delivery is amnesia — the same question answered differently in week one and week five. Decisions get recorded once and stay recorded, which is also why handover documentation isn't a scramble at the end.
03Review
Ugly Sweater
The gate
A visual regression engine built on Playwright. Renders every scenario across a browser matrix, diffs against approved baselines, and captures traces, accessibility findings and Core Web Vitals. Failures come back with an AI-written explanation attached to the pull request.
This is the specific answer to "AI-built means low quality". Unintended visual change blocks the merge. It's the same gate we sell — we run it on this website.
04Ship
Control Tower
The floor
Service management and deployment control plane. Reverse-proxy configuration, staged deploys with promote and rollback, health checks, scheduled backups, and single sign-on across every internal application.
Shipping is a routine operation rather than an event. Rollback is a button, not an incident.
05Build
Matana
The parts bin
A shared React and TypeScript component library — the interface primitives, themes, auth screens and patterns that every application is assembled from.
Standardised inputs. You aren't paying for someone to rebuild a data table, and the tenth screen looks like the first because it came off the same tooling.
06Build · Ship
Snapier
The couplings
A self-hosted integration platform connecting REST APIs through webhook or polling triggers and ordered action chains, with encrypted credential storage.
Most business software fails at the seams between systems. Integration is treated as a first-class component with its own tooling, not glue code written twice.
Throughput
How hard the line runs.
Our own telemetry, March 2026 to August 2026.
31B
Tokens processed
Across all agent work since March 2026.
5,392
Agent sessions
Discrete pieces of work run through the line.
6
Model providers
Routed per task on cost and capability.
US$25.7k
Inference, API-equivalent
What the same work would have cost at list API rates.
Commits
4,108 commits across 91 repositories, last 365 days.
The oldest commit in the portfolio is from February 2026. Nearly all of this was built inside six months.
As at 6 Aug 2026
LessMoreCounted from local git history — most of these repositories are private.
These are figures for our own machinery, not a claim about your project. Delivery numbers — time to ship, cost against a comparable quote, rework rate — get published when there's a named client case study behind them.
Independently listed on
See it run on your problem.
The most useful thing we can do in a first conversation is put your brief through the front of the line and show you what comes out.
Start a build