Artificial Intelligence
The machine writes the repetitive code. We bring the architecture, the data and the judgment, and we sign every line before production. The agent proposes; the engineer disposes.
“I assert that the organizations that build and run agents are responsible for everything those agents do, whether that behavior is intended or emergent.”
“We can outsource many things, but not the acceptance criteria, at some point there’s a human request and a human judgment on whether that request was properly executed.”
We use AI in two ways. As a tool, to build with more quality and fewer hours what we already knew how to build. And as a system component, to solve what was not possible until recently.
When code gets cheap to generate, far more software gets generated (Jevons paradox). And all that new software must be integrated and secured inside systems that already exist. That is the most valuable part of the work, and the part we have done since 2019.
Generated ≠ shipped
An agent is AI that does things, not just answers. To be useful it needs to read the order in SAP, check stock in BigQuery and write the result where your team actually works. That wiring is our home ground.
Our workflow is this very pull request. The agent proposes; the engineer reviews, approves and signs. Always with minimal permissions, every action logged.
git push --force origin main
Agents with bounded permissions. Built on the Agents SDK, they touch only what they must, and a person can always stop them.
Inference at the edge. Document classification and extraction with Workers AI, on the Cloudflare network, in your own account: if Cloudflare already serves your site, you add no new provider.
Governed AI traffic. Every model call logged and cost-capped in AI Gateway. You will know which model answered and what it cost.
Audit before production. Security, business logic and performance. No generated line ships without passing it.
We sign the code
We use AI every day. Claude Code, Cursor, open models on Workers AI. Using it is how we know where it fails.
The industry's most repeated complaint is code that is almost right, but not quite. On a promotion calculation or a stock feed, “almost” is an incident.
So we lead with security. Agents run on least-privilege credentials, work in environments separated from production, and nothing reaches main without human review.
The EU AI Act demands human oversight with the competence and authority to override or halt a high-risk system. That is already how we work.
You won't find an AI product here, because we don't have one. What we have is a method: versioned rules and protocols that steer the agents, and that anyone can audit.
This very website is built that way, by supervised agents, pull request by pull request. And if a chatbot won't solve anything for you, we'll tell you not to build one.
Let's talk, engineer to engineer.
Bring the problem as it is. You'll hear back from the person who'll write the code.
Tell us about your project