THE SAME DISCIPLINE
Security must be built in from the start. We have said that about mobile, about the connected world and about every product we helped to market. AI is the same case. A model call is an untrusted input and an untrusted output at the same time, and it costs money every time it runs.
Since 2026 we build AI into products as a working part of them, not as a demo. The method is the one we used for security: decide what the system may do, measure what it does, and keep the proof in the repository.
WHAT WE DO
Agentic product development
Features where a model reads, decides and acts inside the product: ranked briefings, drafts in many languages, scheduled publishing, advisory sessions with memory. We have shipped all of these in one product, and we know where each one breaks.
LLM architecture and cost control
One router in front of several providers, with a model chosen per feature and every call logged with its tokens and cost. Gemini on Vertex AI, Claude and open-weight models on AWS Bedrock, all pinned to EU regions. We tune thinking budgets per model generation and measure cache behaviour per region before we order a prompt.
AI visibility
How AI assistants describe a brand is now a measurable thing. We measure it by composing the prompts people actually ask and scoring the answers. That is a new kind of search problem, and it is a prompt-composition problem before it is a content problem.
Secure, compliant AI in the EU
Vector memory with per-tenant encryption at rest. The data inventory, the data-processing agreement and the consent registry are written first and the code follows them. Every account signs in with a second factor. Staging cannot leak to search engines or send mail by mistake.
HOW WE WORK
We build with Claude Code, and the build itself is routed like a small department. A planning model writes the specification, briefs the work and reviews every change. Coding models implement one task at a time from that brief. Research models check claims and look for the counter-example.
Nothing goes in without a gate:
- unit tests run before every push, more than 8,000 of them in the current project
- static analysis with a frozen baseline that may only shrink
- a visual regression audit that compares what the page renders with the design tokens it was supposed to use
- a written catalogue of the failures that lint and tests cannot see, each with the test that now catches it
The tooling remembers. Project memory is verified against the git history and the deploy log before a session trusts it, and every plan’s file references are checked before the plan runs. Deploys are delta uploads with a checksum manifest, bootstrap files last, and a drift check afterwards.
The same method rebuilt this website.
A STEALTH PROJECT, IN NUMBERS
Since February 2026, one person with agentic tooling has built a WordPress-based SaaS that is close to release. It is not named here yet.
Inside it sits a multi-model LLM architecture behind one router with a model override per feature, vector memory with tenant-keyed encryption at rest, EU residency for every model call and GDPR handled by design rather than by a banner, and a pipeline that produces a podcast and marketing video without a studio.
It took about 3,500 commits and 290 design documents. The test suite has 640-odd files and more than 8,000 unit tests, and static analysis runs on every push.
One person, with the right method, ships like a department and keeps the evidence.
LATEST ON AI
-
Your AI visibility is a prompt-composition problem
Before you can improve how AI assistants describe your company, you have to know what people ask them. Measuring that is composition work, not content work.
-
Opus plans, Sonnet codes: how a one-person team ships like a department
A planning model writes the brief and reviews the diff. Coding models implement one task at a time. Research models look for the counter-example. That split is the whole trick.
-
Gates that stay green while it is broken
Tests, lint and static analysis each have a blind spot. We keep a written catalogue of the failures that passed every gate, and the test that now catches each one.
Want the right product for the right market with the right timing? Contact us to discuss the possibilities.