About
Built so every finding can be traced.
I’m Stela, founder of FairAudit.
I started Diversalytics in 2019 to help companies comply with Spain’s equality plan and pay transparency rules (RD 901/2020 and 902/2020). Six years of audits taught me one thing above everything else: a finding is only useful if you can trace it back to the data and to the exact provision it comes from.
When the EU AI Act arrived I wanted the same discipline for AI systems. I completed the AI Ethics certification at Turing College on an EU DiversiFAIR scholarship, which led to a scholarship for their AI Engineering Programme and a mentoring role on it. FairAudit is what I built with that: an assistant that answers from the text of the Act, cites the article and recital, records bias tests as evidence, and tells you what it does not cover.
Before all this I trained as a clinical psychologist at Universidad Complutense de Madrid and did doctoral research at the University of Lisbon on how performance reviews in tech treat women differently. That is still the question underneath most of my work.
I’m based near Barcelona. Regulatory information, not legal advice. If you have a hard question about the Act, I’d rather hear that than give you a demo.
Stack: Python, LangChain, LangGraph, Fairlearn, AIF360. FairAudit itself runs Fairlearn only.
The question that shaped the product
A tester asked whether an interview agent that notes a candidate’s tone of voice has a legal problem. It has two. The first version of the assistant missed the prohibition, and the evaluation had scored that answer highly. The fix, and why every citation is now bound to a passage, is written up.
Read the write-up →The method
Four rules, all running in the product today.
One chain, both directions
Obligation, control, test, evidence, report. Every answer cites its article. Every test is logged against a named control. Every report line points back to the obligation it covers.
Code before a model
Never pay a model to verify what a string comparison can verify. Citations are bound to passage ids, so a reference that does not exist cannot be produced.
Evidence that cannot be edited
Each test run is recorded with its dataset, model, metrics and time, in a store that only appends. Reports are assembled from that record.
It signs what it writes
Every output carries a signed, machine-readable declaration. Anyone can verify it against the published key.
The rules, dated
| Regulation | In force | Scope | Maximum penalty |
|---|---|---|---|
| EU AI Act | Phased | Prohibitions since 2 Feb 2025. GPAI duties since 2 Aug 2025. Art. 50 transparency since 2 Aug 2026 (systems already on the market: 2 Dec 2026). Annex III high-risk from 2 Dec 2027 (Digital Omnibus). Annex I product-embedded from 2 Aug 2028. | €35M or 7% of global turnover |
| NYC Local Law 144 | Since 5 Jul 2023 | Bias audits for automated employment decision tools | Up to $1,500 per violation per day |
| Colorado AI Act (SB 24-205) | Since 30 Jun 2026 | Consequential decisions, employment included | State AG enforcement |
| California FEHA regulations | Since 1 Oct 2025 | Automated-decision systems in hiring | Civil liability |
| GDPR Article 22 | Since 25 May 2018 | Automated individual decision-making | €20M or 4% of global turnover |
Statuses as of 29 September 2026. The registry behind FairAudit tracks the dates for each obligation, and RegWatch watches for changes.
FairAudit is built in the EU by Inclusive Data SL. It is grounded in years of academic research and consulting practice, and in a plain aim: make governance accessible enough that teams can show their AI products are responsible.
AI shouldn’t be at the helm of HR. People should, with evidence.
The EU AI Act requires human oversight of high-risk hiring systems (Art. 14). We help you show it is real.
The assistant: sign in with your email. We confirm access by email. The high-risk obligations for hiring systems, Article 14 included, apply from 2 December 2027. Regulatory information, not legal advice.