Work

Reference builds

Four complete systems, written up the way we would hand them to your engineering team: what the work looked like before, what got built, what broke, and what it cost. Plus one open-source piece you can run today.

About these write-ups. These are Levelbrook reference builds — the architecture, guardrails, failure modes and economics of the systems we build, documented as complete engagements. Client names are withheld and the figures are modelled from the operating assumptions stated in each build, not lifted from a named customer’s books. The two live demos are real and run in your browser: the approval queue and the call agent.

Wholesale distribution
14 weeks

A support department that closes its own tickets

Email and chat resolution with real ERP write access, behind a one-click approval queue and an immutable ledger. Six action types graduated to automatic.

→ Home services, 6 sites
7 weeks

The phone that stopped going to voicemail

An AI voice agent on overflow and after-hours, booking straight into dispatch, with warm transfer on any hesitation.

→ Insurance brokerage
9 weeks

Killing 1,900 hours a year of retyping

Certificate and renewal intake: confidence-gated extraction feeding an n8n workflow that writes to the agency management system.

→ B2B healthtech, 60 eng
12 weeks

Rebuilding the development loop around agents

Spec-to-PR drafting, a review agent tuned on the team's own escaped defects, and characterisation tests on a decade-old codebase.

→ Open source, n8n
September 2026

The green run that did nothing

A linter for n8n workflow JSON and an importable hardening kit for the execution that succeeds, writes zero rows, and tells nobody. Not modelled: it is on GitHub, and the numbers came from 400 public templates.