Your company doesn't need more AI. It needs an AI of its own.

Both protect your data, but in different ways. With the agreement, your data travels to the provider's servers and the provider commits not to use it to train its models. With Kelvo, your data doesn't leave: the model, the memory and the documents are on a server that is yours or reserved for you, and every action goes into a log you can review.

The other difference is what happens with what your company teaches. With the agreement, none of it stays in any model: the next day the AI knows as much about you as it did on day one. With Kelvo, what your team approves is used to fine-tune a model that belongs to your company.

The best open models are very close to the best closed ones. In the most widely used public comparison, the gap between the best on each side is 3.3 % (Stanford AI Index, 2026), and on some tests an open model has already beaten the closed ones of its day. The most capable ones are very large and need a lot of compute, and for much of a company's work a smaller one is enough. That's why what matters is choosing the right one for each case: we measure it on your own tasks before recommending one.

On top of that, the chosen model is fine-tuned to your company. Fine-tuning means training it a little further on what your team approves, so it learns your way of working: the structure, the tone and the criteria. That's how it writes like you without being told every time. With closed models this is less and less possible: the big providers are closing it off or limiting it.

Because they do different things, and Kelvo does both.

• Context gives it the data. From day one, Kelvo stores and organizes what happens in your company (documents, systems, meetings), and the AI looks up what it needs, with the exact source.
• Fine-tuning teaches it to work with that data: how your team writes, decides and corrects, and what matters in each case.

What you notice day to day: drafts that come out the way your team would write them, fewer corrections and the same way of working across every case. We start from open models and fine-tune them with LoRA, which trains a small layer on top of the model, on your own server, with what your team approves over time. Kelvo stores that from the start in a format that is ready for fine-tuning.

Since your data is already organized for fine-tuning, the new model is fine-tuned on that same data and tested against the one you have, in two ways: with our test bench, which runs real work cases and saves what each model produces so it can be reviewed, and with your own team, which compares answers from both without knowing which is which and picks the one it prefers.

If the new one wins on both, we propose switching, and we switch only if you approve. In the meantime, the version your team validated doesn't change on its own. That way your AI improves along with the open models, without depending on anyone's terms.

The model, the memory and the documents are on a server that is yours or reserved for you. That doesn't mean working offline: your team's computers connect to the server over the office network or through a private, encrypted connection. What doesn't exist is an open door from the internet into the server.

When the AI searches the web, the only thing that goes out is the query, without your data. Every read, search and written file is logged with the room, the time and whether anything left.

You decide what the whole company knows and what only one group knows:

• The company: the shared memory (templates, criteria, how each thing is done). Whatever goes in there is approved first.
• The rooms: one space per matter, client or area, open only to the people you define. The AI in one room doesn't see another's.
• Personal: each person's own, not shared unless they choose to.

No. Many of these platforms run on models from different AI providers under the hood. We can be one more: we offer your platform our compute and our models, your team keeps using it as before, and the same compute is available to your company.

It can also work the other way around: the platform plugs into Kelvo, and your team works in our system with the platform's skills and industry knowledge.

The law updates personal data protection in Chile. Using AI on servers abroad is lawful with the standard contractual clauses, but it adds an international transfer and more compliance reviews. Processing on your own server avoids that transfer.

In addition, people can ask you for access to their data, to correct it, delete it or take it with them, and you have a deadline in business days to respond. Since in Kelvo the information is organized by room and searchable, it's easier to find and hand over what you hold on someone.

Running locally doesn't exempt you from the rest: the legal basis and the impact assessment are still yours. There is a bill to move the law's effective date to December 1, 2027, so it's worth confirming it when you decide.

It depends on the size of the team and the tasks, and we work it out in the assessment. After that the cost is fixed: you don't pay per question, per document it reads or per fine-tuning run.