Thoughtcoast for AI developers — production AI, not demos

The production surface

What we build.

Four kinds of system, one test: does the number move?

Booking engines

Postcode to confirmed job.

A customer books a real slot in a real diary with no one in the middle — availability, deposit through Stripe, confirmation, done. Built so a retried webhook can never book the same job twice.

Pricing brains

Quotes that show their working.

Rules first, model second. Every figure in an instant quote traces back to a rule a human wrote and approved — and the conversion rate answers for the design, not a slide.

RAG knowledge bases

Retrieval you can score.

Hybrid retrieval — keywords beside vectors, then a rerank — measured against a hand-labelled gold set before anything ships. Recall is a number here, not a feeling.

Back-office automations

The unglamorous middle.

Invoice chasing, CRM hygiene, job handovers between systems that were never meant to talk — Pipedrive, accounts, telephony — joined with n8n and plain REST.

Every one of these runs inside a live business and is measured on its P&L. A system that doesn’t move a line doesn’t survive.

The standards

How we build.

The bar a change has to clear before it ships.

Eval gates before ship. Every retrieval or model change re-runs a hand-labelled gold set in CI. Score below the baseline and the deploy is blocked — by the pipeline, not by policy.
Idempotent everything. A retried webhook can’t create a second job or a second invoice. Delivery is at-least-once; effects are exactly-once.
Fail closed. When an automation stops, its work drops to a human queue. The failure mode is manual and visible — never silent.
Least privilege. The quoting service can read the CRM, not delete from it. Each environment carries its own credentials, and secrets never touch the repository.
One narrator. Every automated change lands in a single audit log — what changed, which system triggered it, when. When two systems disagree, the log settles it.
Humans sign the writes that matter. Money out, records with legal weight: automation drafts, a person commits.

The long version — with a deploy gate you can try to break →

People

Work with us.

Thoughtcoast is small by design. No layers, no ticket mill — a handful of people who each own a system end to end, from the gold set that gates it to the P&L line it answers to.

That shape is deliberate. Production AI inside a live business rewards engineers who can hold the whole problem at once: the model, the plumbing, the failure modes, and the number the system is supposed to move. It is harder than a demo and better than one.

We’re not always hiring, but we’re always talking — engineers, and collaborators who keep engineers honest. If you’ve ever blocked your own deploy because the eval said so, the standards above will read like home. Say hello.

Build the AI a real business runs on.