01
Agent architectures
We specify the graph: tools, memory, handoffs, and the human checkpoints that keep a multi-agent system from improvising.
You leave with a written architecture, not a notebook of prompts.
Lisbon studio · Est. 2023
We design the systems behind the systems.
Nine senior practitioners in Lisbon who architect agents, evaluation harnesses, and retrieval for companies that already know a demo is not a system.
14
Systems in production
9
Specialists in Lisbon
6 mo
Typical retainer
0
Abandoned pilots
01 / Services
What we actually build
We do not sell a platform. We design the four layers that make an AI system operable by your own team.
02
Evaluation harnesses
We build the tests that tell you whether the agent is getting better, not louder.
Golden sets, graders, and regression gates sit in CI so a model swap cannot silently undo a quarter of work.
03
Retrieval pipelines
We design the index, the chunking, the query rewrite, and the failure modes when the corpus is messy.
The point is citations you can defend, not a vector store you cannot explain.
04
Runtime operations
We wire traces, budgets, kill switches, and on-call paths around the model so production is a system, not a hope.
This is the plumbing that lets a nine-person team sleep.
02 / Method
A system before a sprint
Implementation starts after the system is specified. The sequence below is the engagement, not a sales metaphor.
1
Week 0–2
Diagnose
Two weeks in the codebase and the call recordings. We write down what the system actually does, not what the slide deck claims.
2
Week 2–3
Specify
A short architecture note: components, interfaces, evals, and the operational envelope. Nothing is built until this document survives a sceptical principal engineer.
3
Week 3–10
Instrument
We ship the harness first. Agents, retrieval, and tools are then added against failing tests, the same way you would grow a compiler.
4
Week 10–16
Hand over
Runbooks, traces, and a six-week pairing period with your team. We do not leave you holding a system only we can operate.
03 / Work
Three systems in production
Named clients, measured outcomes. Each engagement left a harness and a runbook with the in-house team.
18m → 4m
Median handling time
Meridian Bank
Wholesale operations
Problem
A 40-person ops team pasted between six systems to answer credit questions, and a vendor chatbot hallucinated on 31% of policy answers.
What we built
A tool-using agent over the policy corpus and core banking APIs, with a human checkpoint for exceptions above a stated limit.
Outcome
Median handling time fell from 18 minutes to 4, with citations on every answer and a measured 2.1% error rate.
12% → 67%
Accept-as-draft rate
Vespera Health
Clinical documentation
Problem
Draft notes from a speech model were unusable; clinicians spent longer editing than they had spent dictating.
What we built
A constrained generation pipeline with specialty schemas, retrieval over local guidelines, and an eval set of 1,200 de-identified notes.
Outcome
Accept-as-draft rate rose from 12% to 67% across three specialties in twelve weeks.
41% → 94%
First-response SLA
Atlas Freight
Exception handling
Problem
Delay and customs exceptions arrived as unstructured email; junior coordinators triaged by folklore.
What we built
A retrieval-and-routing agent that classified exceptions, pulled the shipment graph, and drafted the next action for a human dispatcher.
Outcome
First-response SLA recovered from 41% to 94%; the same desk absorbed a 30% volume increase without new headcount.
04 / Studio
Four of the nine
Four of the nine who sign the work. The others are embedded with clients this quarter.
- IC
Inês Carvalho
Founding partner, systems
Previously led applied ML at a Lisbon payments firm. She writes the architecture notes and refuses to ship an agent without a kill switch.
- RA
Rui Almeida
Principal, evaluation
Built the first model-eval platform at a European telco. He treats prompts as untrusted input and graders as the product.
- MK
Marta Kovács
Principal, retrieval
Spent six years on search at a news publisher. She is the person who asks what happens when the corpus is wrong.
- TB
Tomás Berger
Partner, operations
Ex-SRE from a global marketplace. He owns traces, budgets, on-call, and the boring diagrams that keep 3 a.m. quiet.
05 / FAQ
Questions we are asked first
Do you train foundation models?
No. We design the system the model lives in: tools, memory, retrieval, eval, and runtime. If a fine-tune is warranted we will say so and work with whoever trains it.
How small an engagement will you take?
A diagnostic is two weeks and a written architecture. We rarely take implementation under a three-month retainer; the failure mode of short work is theatre.
Will you work inside our stack?
Yes. We adopt your cloud, your repo, and your observability. We are not a platform vendor and we do not leave behind a private runtime you cannot hire for.
Where is the team based?
The studio is in Lisbon, at a workshop in Marvila. The nine of us work remotely with clients in Europe and the US, with overlap hours as a contract, not a slogan.
Who actually does the work?
Two principals on the problem, one partner on the operating envelope, and your engineers paired from week three. We do not staff a dozen juniors to look busy.
06 / Contact
Bring the system, not the pitch deck
Tell us what is in production, what is failing, and what you cannot measure. We reply within three working days with a yes, a no, or a sharper question.
Or write directly to studio@orrery.systems.