AI InfrastructureEnterprisePrivacy-First

Private AI Infrastructure

A fully offline AI assistant for Luxembourg's professional services sector

I researched, designed, and deployed a complete local AI system that runs entirely on company hardware — no cloud, no external API calls, no data leaving the building. Built with Ollama, Llama 3.2, and RAG document retrieval, this project demonstrates my ability to work across AI infrastructure, system architecture, and real-world compliance requirements.

Role

AI Engineer & System Architect

Focus

Local AI, RAG, Privacy Infrastructure

Status

Working proof of concept

OllamaLlama 3.2Open WebUIRAGLinuxAMD GPU
0%
Offline
0
Per token
0 min
Setup time
0 bytes
Data leaked

The Problem

Luxembourg's firms are locked out of AI

Law firms, fiduciary companies, medical practices, and financial advisors handle some of the most sensitive data in Europe. GDPR obligations and professional confidentiality rules make cloud-based AI tools completely off-limits.

ChatGPT, Claude, Copilot — all send data to external servers. For these industries, that's not a risk. It's a disqualification.

ChatGPT / Claude

Data sent to US servers

Microsoft Copilot

Cloud-dependent, GDPR grey area

Fine-tuned cloud models

Still leaves the network

This solution

100% local — nothing leaves the hardware

Architecture

The stack — everything runs on one machine

No cloud dependencies. No API keys. No external calls. The entire system lives on a single Linux workstation with an AMD GPU for local inference.

⚙️
Ollama
Local model runtime
🧠
Llama 3.2
AI language model
🖥️
Open WebUI
Browser interface
📂
RAG
Document retrieval
🐧
Linux
Host operating system
AMD GPU
Local inference

Live Proof of Concept

Same question. Very different answers.

Tested with real Luxembourg legal documents. RAG retrieval grounds every response in verified official sources.

What is a SARL-S minimum capital?

Without documents

€25,000 (incorrect — confidently wrong)

With verified documents

€1 — sourced from Article 720-6 of Loi du 10 août 1915

Tested with: Loi du 10 août 1915 (consolidated 2025) · One-Way NDA template

Deployment

Up and running in under an hour

01

Install Ollama

Single command deploys the local model runtime with AMD GPU support

02

Download Model

Pull Llama 3.2 — 2GB, runs entirely on company hardware

03

Deploy Open WebUI

Browser interface at the company's internal IP — familiar ChatGPT-like experience

04

Build Knowledge Base

Upload verified legal documents, internal procedures, and contracts as RAG sources

05

Train Staff

Lawyers, secretaries, managers — all they need is a browser

Applications

Every department. One tool.

⚖️

Legal

Contract review, clause flagging, GDPR compliance, cross-border document analysis

🏦

Finance & Fiduciary

Regulatory Q&A, client briefings, internal procedure lookup

👩‍💼

HR & Operations

Policy assistant, onboarding guide, job description drafting

✉️

Secretarial

Multilingual email drafting, document summarization, response generation

📋

Compliance

Real-time regulatory cross-referencing against uploaded law documents

🏥

Medical

Patient file summarization, procedure lookup — all patient data stays local

Reflection

What this project taught me

This wasn't a typical web development project. It required me to think at the infrastructure level — understanding GPU drivers, model quantization, network isolation, and how to build retrieval pipelines that produce verifiably accurate results from domain-specific documents.

It also sharpened my ability to identify a real market gap and prototype a solution end to end. Luxembourg has over 1,200 law firms, 400+ fiduciaries, and hundreds of medical practices — all handling confidential data, all currently unable to use AI tools. Understanding that problem and building something that solves it taught me as much about product thinking as it did about technical architecture.

Skills Demonstrated

Local AI model deployment and configuration

RAG pipeline design with document retrieval

Linux system administration and GPU setup

Privacy-first architecture for regulated industries

Understanding of GDPR and professional confidentiality requirements

Bridging technical solutions with real business needs

End-to-end proof of concept — from research to working demo

Outcome

Your data. Your hardware. Your control.

A working proof of concept demonstrating that powerful, private AI is achievable today — with open-source tools, modest hardware, and the right architecture.

This project reflects the kind of work I want to do: finding real problems, understanding the constraints, and building something that actually solves them.

← Back to PortfolioNext: Santiago Voget →