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 engineers building agent architectures, evaluation harnesses, retrieval pipelines, and the operational plumbing that makes AI work in production.
42
systems shipped
6
countries served
100 %
code handed off
What we do
We design multi-agent systems that survive contact with production.
Tool-use graphs, memory hierarchies, routing logic, and failure modes—defined and stress-tested before you write a line of orchestration code.
Off-the-shelf evals measure the wrong things for your domain.
We build automated scoring, human-in-the-loop protocols, and regression suites so you know exactly when your model degrades.
RAG is easy to demo and punishing to operate at scale.
Chunking strategies, re-ranking layers, and hybrid search architectures tuned to your corpus, your latency budget, and your accuracy bar.
The work between the model and the user is where most teams stall.
Guardrails, observability, cost controls, and deployment automation that let AI systems run without hand-holding.
How we work
A two-week diagnostic. We audit your stack, interview your engineers, and document every dependency, bottleneck, and risk surface.
We write the architecture: component boundaries, data contracts, failure modes, and a build sequence. Every trade-off made explicit.
We pair with your team on the hard parts—the orchestrator, the eval suite, the retrieval core—and hand off clean, documented code.
Load testing, red-teaming, runbook writing. We stay until the system runs without us and your team can extend it alone.
Selected work
Client
Challenge
Clinical-notes search returned irrelevant results 40% of the time, slowing diagnosis workflows across twelve hospitals.
What we built
Rebuilt their retrieval pipeline with a hybrid dense-sparse architecture and a domain-specific re-ranker trained on clinician feedback.
Result
94% precision at p95 latency under 200 ms
Client
Challenge
Needed autonomous agents to reroute freight in real time but couldn't trust unverified outputs at scale.
What we built
Designed a three-agent architecture with a verifier loop, confidence scoring, and a human escalation protocol.
Result
6× automated rerouting, exception rate down to 1.2%
Client
Challenge
Their AI tutor hallucinated on 12% of math explanations, eroding trust with teachers and parents.
What we built
Built a domain-grounded evaluation harness with step-level verification and a citation pipeline back to approved curricula.
Result
Hallucination rate fell to 0.3% across 50k sessions
Our team
Principal, Systems Design
Former distributed-systems lead at Feedzai. Thinks in state machines.
Principal, ML Engineering
Built retrieval infrastructure at Spotify. Unreasonably good at failure analysis.
Senior Engineer
Compiler background, now builds evaluation frameworks. Writes the tests nobody else wants to.
Senior Engineer
Operations specialist. Previously ran ML platform reliability at OutSystems.
FAQ
Tell us what you’re working on. We respond within two business days with an honest assessment of whether we can help.