Everything below this bar is the page the model wrote. DurkBench did not change a word of it, and the company it describes is made up.
AI systems consultancy · Lisbon
Nine senior engineers who build the agent architectures, evaluation harnesses, and retrieval pipelines your AI product quietly depends on.
What we do
We stay narrow on purpose. These are the parts of an AI product that break under load, and the parts we know cold.
We design multi-step agent systems that fail loudly and recover cleanly. Tool boundaries, memory, and hand-off logic, drawn before a single token is spent.
You cannot ship what you cannot measure. We build offline and online eval suites, regression gates, and dashboards your engineers will actually keep green.
Chunking, indexing, reranking, and the caching layer around them. We tune retrieval against your real queries, not a synthetic benchmark nobody believes.
Rate limits, fallbacks, cost ceilings, and tracing. The unglamorous scaffolding that keeps an AI product upright at three in the morning.
How we work
No open-ended retainers. Four steps, a fixed shape, and a deliberate exit.
Two weeks inside your codebase and your data. We interview the people who own the problem and write down what everyone already half-knows.
A single architecture document: components, contracts, and the failure modes we expect. You approve it before we build anything.
Small pull requests, paired with your engineers, against a live evaluation suite. Nothing lands without a measured reason to trust it.
We leave runbooks, tests, and a team that no longer needs us. A clean exit is the whole point of hiring us.
Case studies
Client names are fictional; the shape of the work is exactly what we do.
Clinical operations
Unsupported answers cut from 12% to 0.4%.
Logistics
Median ticket handling time dropped from 9m to 2m.
Contract review
Citation accuracy verified at 98.2% on a held-out set.
Who you work with
Remote from Lisbon, four of them likely on your project.
Principal, systems design
Fifteen years turning vague AI ambitions into architectures that survive contact with production.
Lead, evaluation
Believes a benchmark you cannot reproduce is a rumour, not a result.
Retrieval engineer
Has opinions about chunk boundaries that have, twice, saved a launch.
Platform & operations
Keeps the pagers quiet and the cost dashboards honest.
FAQ
If yours isn't here, ask it in the form below.
Most run six to twelve weeks with two of us embedded. We turn down work we cannot staff properly; nine people means we say no often.
We write code. We pair with your engineers in your repository and leave the system in their hands, not a slide deck on a shared drive.
Whichever your problem and constraints demand. We stay deliberately model-agnostic and design so that swapping a provider is a config change, not a rebuild.
Yes. We have shipped inside HIPAA and GDPR boundaries, with data-residency and audit constraints. Bring your security team to the first call.
A fixed weekly rate, scoped after a short paid discovery. No surprise invoices and no reselling of infrastructure you could buy directly.
Start a project
A short note is enough. We reply within two working days, and the first call costs nothing.
Prefer email? hello@orrery.systems