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.
Orrery Systems is a nine-person consultancy in Lisbon. We build the agent architectures, evaluations, and retrieval pipelines that let AI features survive contact with production.
Services
We are not a full-stack agency. Every engagement lands in one of these four bodies of work, usually two at once.
We design multi-step agent systems that hold state, defer to humans, and fail loudly. Built around your domain, not a demo notebook.
Offline suites, replayable traces, and production canaries that catch regressions before your customers do. We treat evals as first-class product.
Chunking, hybrid search, reranking, and freshness policies tuned to the shape of your corpus. Boring plumbing, measurable recall.
Tracing, cost accounting, guardrails, prompt versioning, and rollout tooling. The unglamorous layer that makes the glamorous layer safe to ship.
How we work
We keep the same rhythm across every engagement so your team always knows what happens next.
A short, paid engagement to read your code, interview your team, and write a plan you could hand to anyone. No slide decks.
We build against your real traffic and real edge cases, not a curated dataset. First traces land in your observability stack by day ten.
Evals, red-team runs, latency budgets, and cost ceilings. We ship when the numbers say ship, and we publish the numbers internally.
Your engineers own the system on day one. We leave documentation, runbooks, and an on-call rotation that tapers over the following month.
Case studies
Metrics measured six months after handover.
Team
There are nine of us. These four lead engagements. Every project is staffed with a principal and one or two engineers, no juniors, no proxies.
Fifteen years shipping infrastructure at trading firms and health startups.
Writes the evals nobody wants to write and then makes them fast.
Believes most RAG problems are corpus problems in disguise.
Runs the on-call rotation and the coffee machine, in that order.
Frequently asked
Between six and sixteen weeks, with two to four Orrery people embedded. We turn down work we cannot staff seriously.
Yours. We are comfortable across the major model providers, vector stores, and orchestration frameworks, and we will not sneak a new one in.
Yes. We work under DPAs, BAAs where applicable, and can operate inside your VPC. We do not train on client data, ever.
A tapering support month is included. After that, most clients keep us on a light retainer for evals review and quarterly architecture check-ins.
Selectively. We add roughly one person a year and prefer engineers who have owned a production system end to end. Write to us anyway.
Start a project
A principal reads every inbound within one working day. If we are the wrong fit, we will say so and often point you somewhere better.