A SOFTWARE STUDIO IN MAASTRICHT. SOFTWARE THAT COMES TO YOU.
Most of the work inside a business is a role someone fills by hand: chasing the payments, answering the phone, pulling the numbers for the weekly report. Language models can now do much of that work, but most attempts produce a demo rather than a colleague — something fluent that nobody quite trusts.
We build the other kind. We take one role or workflow and build the system that does it reliably: code holds the facts, the model handles the language, and a person approves anything that matters.
We work remotely from Maastricht with small teams across Europe and the US. We do our best work as the embedded senior on an ambitious team, working directly with the people who own the problem: designing the system, shipping the first version, and bringing everyone else along.
Automating roles with LLMs
The useful agent projects take something a person does every day and do it end to end. That takes a system around the model more than a clever prompt: state the model can't misremember, tools with guardrails, approval on anything that writes, and configurations versioned so each change can be compared with the last. We put an early version in front of the people who'll use it, because the first version teaches you what the job actually is.
Telling whether it's working
Fluent output is cheap and hard to evaluate. We build the measurement alongside the system: a scorecard per measure rather than one pass rate, compared against the previous build, with the raw cases next to every number. The results are often not what you'd guess — on our own systems, changing the output format moved accuracy more than changing the model did. And we'll tell you when the answer is that you don't need this.
Senior full-stack engineering
A decade of React, Next.js, Node, TypeScript, Postgres and AWS — HIPAA-compliant healthcare billing, real-estate platforms, analytics infrastructure. We're at our best alongside product people and stakeholders: quick to prototype, plain about trade-offs, and likely to bring an option you hadn't thought to ask for.
Educational technology
Half of the studio is learning design (M.A. from NYU Steinhardt, elementary-classroom experience). We build software around how people actually learn, and we have a point of view on what AI should and shouldn't do pedagogically — a position the field is still working out, and one we think matters more, not less, as the tools get more capable.
Agents for specialist work
Most of our client work is a version of the same thing: taking a job specialists do by hand and building the system that does it alongside them. Research and reporting, billing and collections, the phone at the front desk — the jobs differ, but the systems rhyme. An agent plans across dozens of tools, queries the company's data lake and researches beyond it, and comes back with analysis or an actual report, as a PDF or slides with charts and images. It keeps a memory of the people, companies and subjects it has worked with, so the next conversation starts where the last one ended. Where money or bookings are involved, it sits on top of a ledger it can read but not rewrite.
What ties it together is a rule we keep coming back to: the model is a clerk, not a scribe. Code holds the facts — the numbers in a report, the balance on an account, the time someone booked — and the model handles the language around them. That's what makes these systems checkable, and we check them. A new ledger runs against every real file from the system it replaces before it goes live. Competing architectures are scored side by side on the same cases. And every change to an agent is measured against the version before it.
Our agent fleet — the agents that run the studio's back office, on local models only. Every write waits for a person's approval. Its memory library is open source as @seneris/memory.
zzppy — bookkeeping for Dutch B.V.s, in its prototype phase. A ledger the database itself keeps honest: every change is logged with who made it and why, and accounting judgment is asked for, never guessed.
DAFTConnect — a networking community for Americans on the DAFT visa. On the App Store.
sleepmixer — a small utility that takes a track apart and lets the pieces decay at different rates. Open source, runs locally.
We're a small studio, deliberately. We move quickly: a working version early, in front of the people who'll use it, before anything gets hardened. Then we measure — what it gets right, what it gets wrong, and what the measurement itself can't see — and write the decisions down so they still make sense six months later. The work we take on has a real customer at the other end and some ambiguity in it, and we don't staff for volume.
We work remotely from the Netherlands and are comfortable in fully distributed, written-first teams. We operate as a Dutch B.V. for European clients. US clients invoice in dollars to a US bank account — the same setup as during the decade we worked from the US.
We offer free or reduced-rate work to animal rescue organisations and bully-breed advocacy groups.
