AI engineering studio
AI systems that
survive contact
with production
Agents that take real actions. Automations that run unattended. Applications built around the problem you actually have. For teams who are past the prototype and need the thing to work on the thousandth run.
- Agents
- That act, not chat
- Pipelines
- That run unattended
- Products
- Live in production
What we do
Three kinds of problem
Most engagements start as one of these and turn out to involve at least two.
Selected work
Systems running in production
Reasonable to wonder
We make trusty agents
“We automated it with AI” is a fair thing to be nervous about. Here is what actually stops an agent we build from doing something you did not ask for.
- 01
MythDeletes your database because it felt like it.
RealityEvery irreversible action (deleting, paying, publishing) sits behind an explicit approval gate. The agent does not get to make that call alone.
- 02
MythEmails your entire list at 3am.
RealitySending windows and per-account rate limits are enforced in code, not in the agent's judgment. It cannot send outside the hours you set, however confident it feels.
- 03
MythConfidently invents a fact and nobody notices.
RealityEvery claim gets checked against a real evaluation set before it ships. Anything the system is not sure about routes to a person instead of guessing.
- 04
MythTurns evil.
RealityIt is a loop that calls scoped tools with a hard iteration limit and a stopping condition. The worst case is it gets stuck, not that it goes looking for trouble.
How we think
Opinions, held for reasons
Four positions that shape every engagement. If you disagree with all four, we are probably not the right studio for you, and that is worth finding out in the first call rather than the third month.
- 01
Evaluation before enthusiasm
We build the test set before the system. If we cannot measure whether it works, we have no business claiming it does, and neither does anyone selling you an AI roadmap.
- 02
Boring where it counts
The model layer moves fast; your data model should not. We keep the novel parts contained and the foundations deliberately conventional, so a model swap is never a rewrite.
- 03
Shipped beats impressive
A narrow thing running in production teaches you more in a week than a broad prototype teaches you in a quarter. We optimise for the former, every time.
- 04
You own everything
Your repository, your cloud accounts, your data, documented for whoever inherits it. There is no layer you have to keep paying us to keep the lights on.
Writing
Notes from the build
Tell us what is slowing you down
A short conversation is usually enough to tell whether this is a build, an automation, or something you should not do at all. We will tell you which.