1. Starting point
Ramirez builds evaluative AI: systems that test documents and processes and flag what does
not hold up. Three rules govern everything we make. Our systems evaluate documents, not
people. They make no decisions; the human decides. Human-in-the-loop sits in the
architecture, not in a disclaimer. And every finding points to its source, so it can be
verified.
2. On this website
This website contains no AI systems you interact with: no chatbot, no automated
decision-making, no profiling. In creating texts and images we use AI as a tool; every
publication is reviewed and approved by a human.
3. In our products
Our systems, including Esmond and Hillmond, evaluate documents against standards calibrated in advance.
They disclose themselves as AI to whoever works with them, issue no scores or verdicts, and
deliver their findings to a human who decides. The classification exercise the EU AI Act
asks of every company, knowing where your AI stands and how it is set up, is one we ran on
our own systems first, and it is built into how those systems are made.
4. Classification per product
The EU AI Act places its heaviest obligations on high-risk AI systems: the uses listed in
Annex III of the regulation. Below, per product, the category that comes closest and where
the product stands. This is our own reading and has not yet been validated by external
counsel. We review it for every new use, every new type of customer and every new guideline
from the European Commission.
Esmond
- Closest category: Annex III, point 3(b): AI systems intended to
evaluate learning outcomes.
- Position in current use: the student uses Esmond on their own
initiative, for their own work. The educational institution does not use it to evaluate
learning outcomes. Esmond gives no grade, no score and no judgement that counts anywhere;
supervisor and jury decide.
- Use by an educational institution: we reassess the classification
before it starts.
Hillmond
- Closest category: Annex III, point 5(a): AI systems used by or on
behalf of public authorities to evaluate whether natural persons are eligible for essential
public assistance benefits and services.
- Position: outside it, for two reasons that each suffice on their own.
The applicant is a company, not a natural person. And support for innovation is not an
essential benefit or service.
- Design: Hillmond gives no score, no ranking and no judgement on
fundability; the assessor decides. Admissibility is not determined by the language model
but by fixed rules in software.
- New modules: we assess each new funding programme separately, in
particular where a natural person can be the applicant.
DownForce
- Position: models and forecasts on business data, such as revenue,
margin and demand. DownForce does not evaluate people; no Annex III category comes close.
We assess each engagement separately.
For all products
- No prohibited practices within the meaning of Article 5.
- Transparency (Article 50): each product discloses itself as AI to whoever works with it.
- Roles: Ramirez BV is the provider of its own systems. The underlying language models
come from providers of general-purpose AI models, including Anthropic.
5. The EU AI Act
The text of the regulation is on EUR-Lex:
Regulation (EU) 2024/1689.
6. Questions
We answer questions about our use of AI, for instance as part of a supplier assessment or a
tender, via
info@downforce.be.