Skip to content
Point of viewSerial number MSK–V/004

We build instead of wrap.Here is why.

Marcos Ruiz · Product engineer, Maistik Studio

On this page
  1. Wrapping is fast. That is the whole argument for it.
  2. Where a wrapper stops being enough.
  3. What we build instead.

Wrapping is fast. That is the whole argument for it.

A wrapper is a thin layer over a model somebody else runs: a prompt, a form, a call to an API. You can have one in a week, and for a demo that is exactly right. It proves that the idea is not absurd. It does not prove that it works on your documents, under your access rules, at the volume your team actually handles.

None of that is a criticism of the model underneath. When the model changes its behaviour, the wrapper changes with it, and nobody at your company gets to decide when.

Most of what is sold as AI today is a wrapper with a logo on it. That is a description of who owns what.

Marcos Ruiz, Product engineer

Where a wrapper stops being enough.

The first wall is data. Your contracts, sensor readings and tickets do not look like the examples the model was trained on. A wrapper cannot fix that; it can only phrase the request more politely. A system can: it decides how documents are split, which fields are extracted, what is retrieved and what is never sent anywhere.

The second wall is evaluation. A wrapper has no way to tell you how often it is wrong on your cases, because it has no idea what your cases are. A system ships with an evaluation set built from your questions, and a score that moves when the model or the data moves.

The third wall is the rules. Who can see which document, what happens when the answer is not in the sources, which decisions need a person to sign. Those rules live in your organisation, and the only place they can be enforced is in software you control.

What we build instead.

We build the four layers as one system: the data layer that prepares and protects what the model sees, the model layer that is chosen and tuned for the task, the agent layer that turns an answer into an action, and the product layer where a person reviews and decides. Compute, a base model or a vector store we buy, because none of them has to be ours. The pieces that encode how your business works, we write.

We buy what does not need to be ours. We build what does.

Marcos Ruiz, Product engineer

That takes longer than a week, and it should. The first version is not where the cost sits. The cost sits in year two, when the wrapper has to be rebuilt because the model behind it changed, or because it never handled the cases that actually mattered.

Network of connections spreading from Valencia across the globe

Working on something *like this*?

Tell us what you are building. We reply with a plain answer on whether AI is the right tool.

Talk to usor write to hello@maistik.studio