01
Agent architecture
We map the decisions your agent actually has to make, then draw the smallest system that makes them reliably. Tool boundaries, memory, escalation paths, and the failure modes you inherit with each.
Lisbon · Nine people · Remote
We design the systems
behind the systems.
Agent architectures, evaluation harnesses, retrieval pipelines, and the operational plumbing that keeps them honest once real traffic arrives.
- Systems shipped since 2021
- 38Systems shipped since 2021
- Median engagement
- 10 wksMedian engagement
- Clients at a time, capped
- 4Clients at a time, capped
- Clients who return
- 91%Clients who return
Trusted by
- Meridian Freight
- Halcyon Health
- Kestrel Ledger
- Ardent Bio
- Northroad Energy
Services
Four things, done properly
We do not staff for breadth. If your problem sits outside these four, we will say so on the first call and point you somewhere better.
02
Evaluation harnesses
Offline suites, graded rubrics, and replay of real production traffic, wired into CI so a regression blocks a merge. Your team keeps the harness after we leave; we write the runbook for it.
03
Retrieval pipelines
Chunking, hybrid ranking, freshness, and permissions treated as one problem rather than four. We instrument recall before we tune anything, so improvements are arguments rather than opinions.
04
Operational plumbing
Tracing, cost ceilings, prompt versioning, rollback, and the on-call rota that keeps a model change from becoming an incident. The unglamorous half that decides whether the rest survives.
How we work
A ten-week arc, written down before we start
Read the system
Week 1
Two of us sit with your traces, tickets, and the engineers who carry the pager. We come back with a written account of where the system actually loses information.
Instrument before building
Weeks 2–3
We stand up measurement first: a baseline eval set drawn from your own traffic, plus tracing you can read. Nothing ships against a metric we cannot defend.
Build in thin slices
Weeks 4–9
One narrow path to production, end to end, then widen it. Our engineers work inside your repository and your review process, not alongside them.
Hand over and leave
Week 10
Runbooks, an architecture note your next hire can read, and two weeks of paired on-call. We schedule the exit at the start of the engagement, not the end.
Case studies
Three systems, and what changed
Meridian Freight
Logistics · Rotterdam
- Problem
- A customs-classification agent was right often enough to be trusted and wrong often enough to be expensive. Nobody could say which cases it failed on.
- What we built
- A graded eval harness over 4,200 historical declarations, a retrieval layer scoped to tariff revisions, and a confidence gate that routes ambiguous goods to a human broker.
- Outcome
- Misclassification rate fell from 6.1% to 0.9% across two quarters, with 31% of cases escalated.
6.1% → 0.9%
Halcyon Health
Clinical operations · Dublin
- Problem
- Clinicians abandoned an internal answer tool because it cited the wrong protocol version. Trust, once lost, did not return on its own.
- What we built
- A permission-aware retrieval pipeline with document-level effective dates, plus a citation renderer that shows the exact revision and refuses to answer when sources conflict.
- Outcome
- Weekly active clinicians went from 40 to 610 in five months; unsupported answers dropped below 1%.
40 → 610 users
Kestrel Ledger
Fintech · São Paulo
- Problem
- A reconciliation agent cost more per run than the analyst it replaced, and no one could attribute the spend to a step.
- What we built
- Span-level cost tracing, a cached planner, and a rewritten tool layer that batches ledger reads. We deleted two of the six agents entirely.
- Outcome
- Cost per reconciliation run fell 74%; median latency went from 96 seconds to 22.
−74% cost
Team
Nine of us. You get the four who fit.
No account layer, no juniors billed as seniors. The people on the first call are the people in your repository.
Inês Corvalho
Principal, systems design
Ten years of distributed systems before any of this had a name; she draws the architecture and defends it.
Tobias Reinhardt
Lead, evaluation
Ex-search quality. He believes an unmeasured improvement is a rumour and will say so in your standup.
Amara Osei
Lead, retrieval
Built ranking infrastructure for a legal archive of nine million documents; now she does it in ten weeks.
Rui Baptista
Principal, operations
Carries the pager on every engagement so your team learns what carrying it should feel like.
FAQ
Questions we get on the first call
How big is an engagement?
Most run eight to twelve weeks with two or three of us embedded. We take on four clients at a time; that is a hard ceiling, not a sales tactic.
Do you write production code or just advise?
We write code, in your repository, reviewed by your engineers. An architecture document nobody can implement is a failed engagement, and we have written a few.
Which models and vendors do you work with?
Whichever survives your evals. We build the abstraction that lets you swap providers, then argue for the boring choice that keeps working.
What if we do not have evaluation data yet?
Almost nobody does. Week two is spent turning your logs, support tickets, and expert judgement into a graded set of a few hundred cases you actually trust.
Can you work with our security constraints?
Yes. We have run engagements inside VPC-only environments and on air-gapped hardware in Frankfurt. Send us the requirements before the call and we will tell you what changes.
Start here
Tell us what is breaking in production.
Write a paragraph, not a brief. Inês or Rui replies within two working days, and the first conversation is an hour of ours, free.
- Next intake
- One slot open, October 2026