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.
Lisbon · RemoteNine people, one discipline
Agent architectures, evaluation harnesses, retrieval pipelines, and the operational plumbing that keeps them honest in production.
01Services
We turned down the fifth. Depth in a narrow band is the whole point of a nine-person firm.
We design multi-step agents with explicit state, bounded tool access, and failure paths you can read. You get a system diagram your engineers will still agree with six months later.
We build the test suites that tell you whether a model change is an improvement or a regression. Graded datasets, judge calibration, and dashboards wired into your CI.
Chunking, indexing, reranking, and freshness are decisions, not defaults. We tune each against your real queries and leave behind the tooling to re-tune them.
Tracing, cost attribution, prompt versioning, rollback, and on-call runbooks. The unglamorous layer that decides whether a demo becomes a product.
02How we work
Two weeks reading your code, your traces, and your incident history. We end with a written assessment naming the three things most likely to break.
A reference architecture, an evaluation plan, and a cost model. Reviewed with your team in one room until nobody has an unanswered objection.
We pair with your engineers rather than shipping a black box. Every component arrives with tests, a runbook, and a named owner on your side.
We stay on a light retainer for one quarter, then step back. If you still need us daily after that, we designed it wrong.
03Case studies
Every number below was measured by the client, not by us, on production traffic.
Logistics · Rotterdam
−94%
Identifier errors in production
Retail · Manchester
71%
Conversations resolved without handoff
Clinical software · Porto
3 wk → 4 h
Time to validate a model change
04Team
The other five are engineers who would rather be in your repository than on this page. You will meet them in week one.
Founding partner, architecture
Spent nine years building trading infrastructure before deciding agents were the more interesting distributed system.
Partner, evaluation
Former measurement lead at a search company. Believes every metric should come with a story about how it can lie.
Principal engineer, retrieval
Has rebuilt the same document pipeline four times for four industries and still finds chunking strategy genuinely contentious.
Principal engineer, operations
Ran on-call for a payments platform. Writes runbooks first and code second, to everyone's mild irritation.
05FAQ
Just the AI systems and the operational layer around them. We work inside your codebase alongside your team; we are not a product studio and will say so early if that is what you need.
All the major ones, plus open-weight models you host yourself. We are deliberately unaffiliated. Our architectures assume you will swap providers at least once, because every client so far has.
Diagnosis takes two weeks and is priced fixed. Design and build usually run eight to sixteen weeks. The handover retainer is one quarter. We decline engagements that look open-ended.
No. We work inside your cloud account with access you grant and can revoke. Evaluation datasets stay with you. We sign your DPA, not the other way round.
Diagnosis is €18,000. Build phases are scoped weekly at a published day rate, agreed before we start. We do not bill for rework caused by our own design mistakes.
06Contact
We reply within two working days, usually with three questions and a suggested time. We take on two new diagnoses a month.
Prefer email? hello@orrery.systems