Closing Brackets
Start a project

What we build, and how a project runs

Every engagement follows the same five stages, and you can tell which one you are in at any point. Below are builds that show what comes out the other end.

Evergreen Softwash — launch film still
Concept buildExterior cleaning · residential services

Evergreen Softwash

A regional soft-wash company with a full crew calendar and no way to fill the gaps. Every booking came through a phone number one person answered, and every quote meant driving out to the property. We built the whole front end of the business: the film that gets attention, the site that converts it, and the agents that answer at two in the morning.

What was built

  • Brand identity and a launch film, cut as a single dusk hero spot
  • Booking site with instant quoting from address and roof area
  • Lead qualifier agent that books straight into the crew calendar
  • Review responder that answers every review inside an hour
  • Local SEO and Google Business Profile across four service areas

What it shows

What the full stack looks like on one small business: film, site, and three agents running the front of house, bought as one scope instead of from five vendors.

  • Next.js
  • Stripe
  • Twilio
  • Claude agents
  • GBP API

Lines of work involved

How the film was cut
Concept buildRegional logistics · 40 trucks

Halcyon Freight

Rate confirmations, bills of lading and proof-of-delivery photos arrived as email attachments and got typed into a TMS by hand. Roughly nine hundred documents a week, three people doing it, and a two-day lag before anyone could invoice.

What was built

  • Document extraction agent reading rate cons, BOLs and POD photos
  • Confidence scoring, with anything under threshold routed to a human queue
  • Two-way sync into the existing TMS — no migration, no replacement
  • Exception dashboard showing what the agent could not read, and why

What it shows

The pattern we use for document-heavy back offices: never replace the system of record. Sit beside it, and hand back only what you are sure of.

  • Python
  • Claude
  • Postgres
  • TMS API

Lines of work involved

Concept buildHealthcare · six locations

Meridian Dental Group

Six practices, six front desks, and a recall list nobody had time to work. Patients who lapsed at eighteen months were never called, because calling them was always less urgent than the person standing at the counter.

What was built

  • Recall agent working the lapsed list by SMS, per location and per clinician
  • Intake forms that pre-fill from the practice management system
  • Escalation to a named human the moment a message reads as clinical
  • Per-location reporting the regional manager actually opens

What it shows

How we scope AI in a regulated setting: the agent handles scheduling language only, and every clinical thread goes to a person by design rather than by exception.

  • Next.js
  • Claude
  • Twilio
  • PMS integration

Lines of work involved

Concept buildDTC e-commerce · homewares

Foundry & Fern

Good products, good photography, and a storefront that took nine seconds to become interactive on a phone. Paid traffic was being bought and then lost somewhere on the way to the product page.

What was built

  • Storefront rebuild against the existing catalogue and checkout
  • Core Web Vitals work: 9.1s to 1.4s time-to-interactive on a mid-range Android
  • Abandoned-cart sequence with per-customer copy rather than one template
  • Attribution wired so paid spend maps to revenue, not to sessions

What it shows

That performance work is growth work. Nothing in the funnel changed except how fast it arrived.

  • Next.js
  • Shopify Storefront API
  • Klaviyo
  • GA4

Lines of work involved

Concept buildB2B SaaS · $18M ARR

Cartwright & Vale

Churn was visible in the numbers about a month after it was decided in the product. Support tickets, usage decay, invoice disputes and quiet champions leaving all lived in different systems, and nobody could see them as one shape.

What was built

  • Account graph joining product usage, support, billing and CRM into one model
  • Signal agents watching each edge for the patterns that precede a cancellation
  • Weekly risk digest to customer success, with the evidence trail attached
  • A written record of every signal that fired and what happened next

What it shows

The graph-first approach we write about: model the relationships first, and questions you could not previously ask become ordinary queries.

  • Postgres
  • dbt
  • Claude agents
  • Metabase

Lines of work involved

The process

How a project runs

The same five stages every time, in the same order, so you can always say which one you are in and what comes next.

  1. Scope

    We work out what you actually need and write it down. You get one document: what gets built, a timeline with real dates, and one fixed price.

  2. Build

    Work lands in stages, and every stage comes off the floor as software you can use. You see progress the week it happens rather than at a reveal.

  3. Integrate

    Where AI belongs in the build, it goes into your existing systems and runs against your data, your workflows, and your business rules. Pilot first, then production.

  4. Hand over

    You own all of the code. Nothing is locked to us, and nothing about the handover depends on keeping us on retainer.

  5. Grow

    If you want the marketing side too, it becomes a machine we keep running, bringing in customers every month while we tune what performs best.

Stage one

Start with a fixed-scope plan

The first stage is the scope document. Tell us what you want built and you will have what gets made, a timeline with real dates, and one fixed price — before any work begins.

Get a fixed-scope planSee what we do