Back to home

Compliance and AI Published tool 2026

compliance-uy

Data protection law, turned into an audit of your repository.

Where it came from

On every project where I handled personal data I found the same scene: the regulation exists, it's public, and still nobody on the technical team knows what it means for the code they're writing. The gap between the law and the implementation is where compliance gets lost.

The problem

Complying with Uruguay's Law 18.331 requires knowing what personal data your system processes, which providers you share it with and where it leaves the country. That information is in the code, scattered across integrations and data models, but the assessment happens in meetings and ends up in a document that looks nothing like the real system.

The solution

An open source skill for Claude Code that reads the repository and produces an initial compliance package for Uruguay. It detects personal data and providers in the code, maps international transfers, evaluates fourteen controls and produces documents and an auditable state where every finding points to a specific file and line.

What's solved

  • Grounded in official sources: Law 18.331, Decree 414/009, arts. 37-40 of Law 19.670, Decree 64/020 and guidance from the URCDP (Uruguay’s data protection authority).
  • Every regulatory claim cites the rule, article and source file in /sources.
  • Anything that can't be verified is explicitly marked [verificar contra fuente oficial] instead of being filled in by guesswork.
  • Detects personal data: email, phone, national ID, tax ID, address, IP, location, messages, files, payments, health, minors and biometrics.
  • Detects more than twenty providers (AWS, GCP, Azure, Supabase, Stripe, Mercado Pago, OpenAI, Anthropic, Twilio, Datadog and others) and maps international transfers.
  • Evaluates fourteen controls and produces a state.json with status, file:line evidence, risk level and suggested remediation.

Evidence

  • Rewritten from scratch for Uruguay: it contains no legal text from other jurisdictions.
  • Explicit warning in the README: it produces drafts and technical assessments, not legal advice.
  • Published as open source, installable in any repository.

The hard parts

  • A tool that talks about law can do harm if it sounds more certain than it is. Almost all of the design went into deciding what it asserts, what it flags as pending verification and how it cites it.
  • Detecting personal data in real code is noisy: an email field may belong to a user or to an internal log. The rule became to flag and ask for confirmation, never to conclude on its own.

What I learned

  • An agent is most useful when its output is verifiable. The rule of citing the rule, the article and the file:line is what makes the result reviewable.
  • Translating regulation into concrete controls is product design: the hard part isn't reading the law, it's deciding what question to ask the code.

Next steps

  • Extend coverage to other jurisdictions in the region while keeping strict per-country separation.

Applied learnings

Reading that grew out of this project