Legibus

The problem

In law, a wrong answer is not an approximation: it is malpractice. Consumer AI assistants answer with confidence, cite nothing, and ship their users’ data to third parties. Unusable for a professional bound by confidentiality.

What I designed

Legibus starts from the opposite principle: the AI may only answer what it can prove.

• A RAG pipeline where ingestion, indexing and generation are strictly separated
• LLMs orchestrated in a controlled chain, no black box
• Versioned embeddings and data, so an answer can be explained after the fact
• A Django platform with its control APIs
• Isolated, sovereign infrastructure

My role

I act as technical architect, AI integrator and infrastructure owner, working alongside a PhD in AI and an R&D team. I own the architecture decisions, the backend and operations.

What it proves

• You can build serious AI without OpenAI: the model is only a small part of the project
• Document quality and evaluation matter more than the choice of LLM
• Explainable AI is designed into the architecture, not bolted on afterwards

AI should not replace. It should shed light.