The question before the model
Most AI work fails in the framing, not the training. We spend the first weeks establishing whether the problem has an answer in your data at all — and we will tell you when it does not.
Studio
Judgement before machinery.
Founded in 2021, Neuryn is a small team of engineers and designers shipping applied AI into contexts where being wrong has a cost.
We started Neuryn because the interesting part of applied AI was being skipped. Teams were arriving with a model already chosen and a problem still undefined, and the work that mattered — deciding what correct meant, in writing, with names attached — kept getting deferred until production made it urgent.
So we do that part first. A Neuryn engagement opens with two to three weeks of discovery that ends in a written recommendation, and roughly one in five of those recommendations is that the project should not proceed as scoped. We would rather lose the build than deliver a system nobody can defend.
What we do build tends to stay built. The oldest system we shipped has been in production for three years without our involvement, which is the outcome we are actually optimising for.
Principles
Four things we will not trade away.
Measured or it did not happen
Every system ships with the evaluation suite that proves it works. Aggregate accuracy is not a result; a reproducible set of cases with a cost and latency budget is.
Built to be handed over
We pair with your engineers from week one and plan the handover before we plan the architecture. If your team cannot maintain it after we leave, we failed.
No lock-in by construction
Providers change, prices change, capabilities change. We build the abstraction that lets you switch without a rewrite, and we benchmark on your data rather than on a vendor benchmark.
Track
How we got here.
- 2021 Neuryn starts as two people and one retrieval problem nobody else wanted.
- 2023 First agent system into production in a regulated environment. It is still running.
- 2024 The evaluation practice becomes the first thing we sell rather than the last thing we build.
- 2026 Twelve clients, thirty systems, one team that has never exceeded eight people.
Team
The people doing the work.
We have never exceeded eight people, and the people you meet in the first call are the ones who write the code.
DESIGN
Mira Lindqvist
Design lead
Designs the interfaces that sit on top of the models — the part that decides whether people trust the output or quietly stop using it.
ENG_LEAD
Nils Aubert
Founder & principal engineer
Builds the retrieval and agent infrastructure behind Neuryn's engagements. Ten years in distributed systems before applied AI, which mostly taught him how expensive a bad abstraction gets.
Bring us the problem, not the solution.
The engagements that go best start with a description of what is broken rather than a specification of what to build.