1. Home
  2. Artificial intelligence

Artificial intelligence

AI in real products, with clear limits.

We build language models into products: assistants that answer from your documents and cite the source, agents that work with a human in the loop, generated content that is checked automatically. With cost caps enforced in code and evaluation before launch.

01 — What we build

AI that finishes the job.

The hard part is not the model. It is everything around it: what data it gets, what it is allowed to do, what it costs and how you check what it wrote. That is what we build.

  1. 01

    Language models inside your product

    We build LLMs into the application, in the context where people actually work: on their page, with their data, within what they are allowed to do. Not a chat box bolted onto the side.

    • LLM
    • In-context assistant
    • API

    Live in the AI Professor in Cursuri AI

  2. 02

    Assistants on your documents (RAG)

    Semantic search across your company documents, with answers that cite their source so anyone can check them.

    • RAG
    • Semantic search
    • Cited sources
  3. 03

    Agents with a human in the loop

    Workflows that read, propose and prepare the actions. Whatever matters is approved by a person before it happens.

    • Agents
    • Human approval
    • Automation
  4. 04

    Generated content, checked automatically

    Generated text goes through automated checks (structure, rules, forbidden wording) before it reaches people. We build case generators the same way.

    • Generation
    • Automated checks
    • Case generators
  5. 05

    Cost caps enforced in code

    Every AI feature has a budget written into the application: a call that would exceed it never goes out. Costs stay predictable.

    • Budget per feature
    • Cap in code
  6. 06

    Evaluation before launch

    The model's answers are tested against cases prepared in advance, before users ever see them.

    • Evaluation sets
    • Tests
  7. 07

    The transparency the AI Act requires

    Users know when they are talking to an AI system and what its limits are. We write these notices straight into the product.

    • AI Act
    • Transparency

02 — How a model thinks

One word at a time, paying attention.

A language model writes one piece of text at a time: a token. For every new token it weighs how much each earlier word matters. That is attention: the arcs you see, thicker and brighter where the link is stronger.

The glass tiles
the context: the question and the instructions
The gold tiles
the tokens the model is writing right now
The arcs
attention: how much each word matters for the new token
The cloud below
your documents, as points in a space of meaning (embeddings)

03 — Proof: the AI Professor

We use it every day, in Cursuri AI.

The AI Professor is built into every lesson of Cursuri AI, our platform of artificial intelligence courses. You type or speak. It is included in the subscription.

  • Explains a passage of the lesson
  • Gives a real-world example
  • Summarises the lesson
  • Generates a quiz to check yourself
  • Makes flashcards
  • Has a fair-use limit

The AI Professor is an artificial intelligence system, not a human teacher; its answers may contain errors.

04 — Startup programs

Members of two startup programs.

Cursuri AI, our platform, is a member of Claude for Startups, Anthropic's program for startups, and of AWS Activate. The first brings regular sessions with Anthropic's Applied AI team; the second comes from AWS, where the Cursuri AI infrastructure runs.

Claude for Startups Members since 2026

Regular sessions with Anthropic's Applied AI team.

A member of Claude for Startups, Anthropic's program for startups; regular sessions with Anthropic's Applied AI team. Their experience helps us develop the product and keep it safe.

About the program, on claude.com ↗
AWS Activate What we build on

Data kept in the European Union.

Infrastructure
Cursuri AI runs on AWS.
Data
In Frankfurt, with a daily copy in Ireland.
Review
Cloudvisor, an AWS Advanced Tier partner, reviewed access rights, secrets and code delivery.
Partnership
We are Cloudvisor partners: we can help you too with AWS credits, optimisation and an AI readiness assessment of your infrastructure.

AWS, AWS Activate, Claude and Anthropic are trademarks of their respective owners. Membership of a startup program does not mean that Amazon or Anthropic recommends our products.

05 — How we work

An AI project, step by step.

We start from the problem to solve, not from the model. The model comes later, and it can change.

  1. 01

    The use case

    What the AI has to do, for whom and, just as important, what it must not do.

  2. 02

    Data and sources

    Which documents and data the model receives and where they come from, so its answers can be checked.

  3. 03

    Prototype and evaluation

    A prototype inside the application and a set of test cases we run on every change.

  4. 04

    Caps and checks

    The budget enforced in code, the automated checks and the points where a person decides.

  5. 05

    Launch and follow-up

    Launch with the transparency notices in the product, then we track cost and answer quality.

Questions

Questions about AI.

Which language model do you use?

We choose the model by task, cost and data requirements, and we build the application so that it stays the same when the model changes.

What does Claude for Startups mean for you?

Cursuri AI has been a member of Claude for Startups, Anthropic's program for startups, since 2026: regular sessions with Anthropic's Applied AI team, whose experience helps us develop the product and keep it safe. Membership does not mean that Anthropic recommends our products; Claude and Anthropic are trademarks of their respective owners.

Where is the data kept?

For Cursuri AI, the infrastructure runs on AWS with the data in the European Union: in Frankfurt, with a daily copy in Ireland. On a new project, where the data lives is one of the first questions we settle.

Can a language model get things wrong?

Yes. A language model can make mistakes, which is why we build with cited sources, automated checks and, where it matters, a person who approves. Users are told they are talking to an AI system and that its answers may contain errors.

How do you keep costs under control?

With cost caps enforced in code: every AI feature has a budget, and a call that would exceed it never goes out. We keep tracking cost after launch.

What does RAG mean?

Retrieval-Augmented Generation: before answering, the application searches your documents for the right passages (semantic search) and hands them to the model. The answer rests on them and cites them, so it can be checked.

What does the EU AI Act require from a product with AI?

Among other things, transparency: people must know when they are interacting with an AI system. We write these notices straight into the product, as we do for the AI Professor in Cursuri AI.

Next step

Got an idea for AI? Let's see what it can do.

Tell us what you want to automate or which questions should get an answer; we will write back with the next steps.