Agents · RAG · Evals · MCP
Hire AI and LLM engineers
Most LLM projects do not fail on model choice. They fail because nobody built the eval that would have caught the regression, or because the retrieval layer was never the bottleneck people assumed it was. We build the unglamorous parts that decide whether an AI feature survives contact with real inputs.
We work your hours
Florianópolis, Brazil, UTC−3 — one to two hours ahead of US Eastern, four to five ahead of Pacific. A shared workday, not a handoff. How nearshore works.
- San Francisco
- your morning
- New York
- your desk
- Florianópolis
- us, working
What we build
- Agent architecture and context engineering
- Agents designed around what goes into the context window and why. This is usually the highest-leverage work available, and it is almost always cheaper than reaching for a larger model.
- Retrieval and RAG pipelines
- Retrieval that actually retrieves, with the parallelization and caching to make it fast enough to sit in a user-facing path.
- Evals and benchmarks
- The thing that tells you a change made the system worse before your users do. Built against your task, not a public leaderboard.
- MCP servers and tool integrations
- Giving models real access to your systems through typed, bounded tools rather than a prompt that describes an API and hopes.
- Cost and latency work
- Routing, caching and model selection tuned so the feature is affordable at your actual traffic rather than at demo traffic.
- Python
- Claude Agent SDK
- MCP
- FastAPI
- Vector search
- Terraform
Proof
Cosmos Reader
An AI reading platform whose agents scored poorly on the company's own benchmark. We rebuilt them around disciplined context engineering rather than a bigger model, and lifted the score to roughly 90%. Separately, for Tavern Research, parallelized retrieval made RAG operations run 50× faster.
50×
faster retrieval · ~90% on a client's benchmark
Start small, expand on evidence.
No twelve-week discovery phase. The fastest way to find out whether this works is to put something in production.
- 01
A call, not a questionnaire
Thirty minutes on what you are building and what is in the way. If we are not the right shop, we will say so on that call.
- 02
A scoped first slice
One real deliverable with a defined edge — an eval suite, a storefront section, a pipeline that was too slow. Small enough to judge, big enough to matter.
- 03
Embedded delivery
In your tools, your review process, your standups. Because we work your hours, review cycles close the same day.
- 04
Expand or stop, on evidence
The first slice tells you what we are worth. Scale up, or walk away having spent very little finding out.
Hire AI and LLM engineers: common questions
Are you tied to a particular model provider?
No, and we build so you aren't either. The routing and eval layers sit above the model so it can be swapped when a better or cheaper one ships — which, at the current pace, it will before your project is a year old.
Our AI feature works in demos but not in production. Can you fix that?
That is the most common engagement we take. The gap is almost always inputs: demos use clean, representative examples and production does not. The fix starts with an eval built from real failures, because without one every change is a guess about whether things improved.
Do you build agents or just integrate an API?
Both, and the honest answer is that a lot of work labelled 'agentic' should be a single well-constructed call. We will say when a simpler design gets the same result, because multi-step agents cost more, fail in more ways, and are harder to evaluate.
Can you work with our data privacy constraints?
Yes. We have built GenAI systems inside enterprise client teams at BCG and under NDA elsewhere, so working within a client's data boundaries and review process is the normal condition rather than a special case.