aeternum
Let's talk

02 / what we do

AI development that ships as software, not as a demo.

Retrieval, agents and model integration, built with the same tests, interfaces and error handling as the rest of your product.

what's included

A model is a component, not a product. We test it like one.

  • RAG & Retrieval
  • Agents & Automation
  • LLM Integration
  • Evaluation

What AI engineering means here

An AI feature is ordinary software with a probabilistic component in the middle. It needs a schema, error handling, retries and a way to tell whether it got worse last Tuesday. The last one is where most LLM projects come apart, so we build the evaluation harness before we build the feature.

Aeternum treats a model the way we treat a payment provider: a dependency with a contract, a failure mode, and a test suite pointed at it.

What we build

  • Retrieval (RAG). Document ingestion, chunking, embeddings and a retrieval layer you can inspect. A wrong answer is traceable to the chunk that caused it.
  • Agents and automation. Tool-calling workflows with explicit boundaries: what the model may do on its own, what needs a person, and what happens when a tool call fails.
  • LLM integration. Model calls behind a single interface, with the prompt, the output schema and the retry policy in version control next to everything else.
  • Evaluation. A graded test set that runs in CI, so quality is a number you can watch rather than an impression.

Why the evaluation comes first

A model upgrade will improve nine of your prompts and quietly break the tenth. Without a scored test set there is no way to know which one, and no way to argue about it with a client.

We collect the eval set from real inputs before any prompt is written. It is the difference between a feature you can change and a feature nobody dares touch.

What we will talk you out of

Chat interfaces bolted onto products that do not need them. If the underlying task is a form with six fields, a form with six fields is faster for the user and cheaper for you. We would rather build the small thing that works.

how a build runs

How a build runs

  1. Find the task, not the model

    We start from the job a person is doing today and how long it takes them. If that job has no measurable output, an LLM cannot improve it and we will say so before you spend anything.

  2. Write the evaluation set first

    A fixed set of real inputs with graded expected outputs, written before the first prompt. Every prompt and model change is scored against them from that point on.

  3. Build the feature around the call

    Typed interfaces, retries, timeouts, fallbacks and a logged trace for every request. The model call is one component in the middle of ordinary software.

  4. Ship with the eval in CI

    The test set runs on every pull request. A regression fails a build instead of surfacing in a client demo three weeks later.

questions

Questions about AI Engineering.

What is RAG, in one sentence?

Retrieval-augmented generation is the pattern where a system searches your own documents for relevant passages and passes them to a language model as context, so the answer is grounded in your content rather than in the model's training data.

Which models do you build on?

That is a per-project decision and we revisit it, because the ranking changes every few months and a choice made once is a choice going stale. Every model call sits behind a single interface with its prompt and output schema beside it, so changing provider is a configuration change rather than a rewrite.

How do you stop the model making things up?

Three things, in order of effect. Ground every answer in retrieved passages and cite them, constrain the output to a schema the code validates, and score the whole pipeline against a fixed test set so a regression is visible. Hallucination is a measurable rate, not a yes or no.

Where does the client's data go?

Wherever you tell us it may go, and that is settled before any code is written. We map which data would leave your systems, what the provider terms say about retaining it, and whether the job can instead be done by a model running inside infrastructure you already control. If the answer changes the design, week one is when you want to know.

posts on this

Posts on this

  1. The AI Act got delayed. Three parts of it did not.AI4 min read
  2. The Swiss case for running a model on your own hardwareAI6 min read
All posts

Tell us what you want to build.

Send the brief. You get a scope and a range within two working days.

other services

Other services