Skip to content
View SFHAJJI's full-sized avatar

Highlights

  • Pro

Block or report SFHAJJI

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
SFHAJJI/README.md

Soufien Hajji — AI & Data Architecture

I build production systems where AI, data engineering, cloud architecture and verifiable evidence meet. Based in Luxembourg; working in English and French.

I am open to AI Engineer and AI/Data Solution Architect roles.

Flagship work

Project Problem solved Engineering evidence
Lex · case study · source Read and compare what Luxembourg and EU law said on a chosen date, article by article, with the official source and a permalink. Formex and Akoma Ntoso parsed into relational SQLite indexes; deterministic question parsing; structure-first retrieval with no vector store; typed answers and twenty typed refusals instead of generated text; machine gates with shuffled controls; reproducible, signed releases on Azure Container Apps; a public MCP server.
Assistant du guide « Vivre au Luxembourg » · source Answer a newcomer's question inside a public guide to Luxembourg: the assistant recognises the fiche being read, opens the right fiche when the answer is elsewhere, and highlights the passage that carries it. Four bricks over the guide's own knowledge base (fiche, section, bloc, FAQ tables); typed question with an expert dictionary in four languages; retrieval as filtering with a margin-based arbiter; a typed answer validated against the guide before display; absence justified with four kinds of evidence; evaluation per question type; the model is optional on every path.
Energy Copilot · source Turn public geospatial, tender, tariff and legal data into traceable solar-prospecting decisions. Azure agentic RAG with deterministic tools, PostGIS, source citations, security controls, observability and cost controls. The public deployment is offline; the architecture page documents it.

The emphasis is not a chatbot demo. It is the work that makes an AI product defensible: temporal data models, retrieval quality, provenance, typed failure modes, deployment safety, operational cost and honest limitations. Lex's assistant is built to a typed answer contract (the model only plans; code validates and quotes) and is switched off until its gates pass.

Try the public, read-only Lex MCP server without a key:

claude mcp add --transport http lex https://law.soufien.lu/mcp

Start here

I welcome technically serious feedback, legal-domain review and conversations about responsible AI systems in regulated environments.

Popular repositories Loading

  1. lex lex Public

    Point-in-time Luxembourg and reviewed EU law: exact article history, deterministic keyword search, gated local hybrid retrieval, signed artifacts, and read-only MCP tools.

    C# 9

  2. lex-ops lex-ops Public

    Lex fleet operations: official-publisher ingestion, signed whole-artifact releases, GitHub OIDC deployment, and three-state freshness records.

    Shell 6

  3. lex-articles lex-articles Public

    Machine-readable Luxembourg and reviewed EU law: per-article Markdown and JSON, validity intervals, and hashes chained to official publisher evidence.

    6

  4. lex-corpus-lu-legilux lex-corpus-lu-legilux Public

    Every consolidated version of Luxembourg law as verbatim publisher files + dated JSON metadata, append-only, sha256-hashed. The tree is the legislative history.

    5

  5. lex-corpus-eu-eurlex lex-corpus-eu-eurlex Public

    Reviewed temporal EU law from EUR-Lex and Cellar: official English and French expressions, consolidations, relationships, hashes, and append-only observations.

    HTML 5

  6. SFHAJJI SFHAJJI Public

    AI and data architecture portfolio: temporal legal search, Azure agentic RAG, and production engineering.

    5