Home / Projects / Novacreux
Final Degree Project · ASIX 2025-2026

NOVACREUX

Modular technology ecosystem that brings generative AI, automation and security to 5 different hardware platforms: a self-hosted homelab, a PC with a dedicated GPU, a Raspberry Pi with an AI accelerator, an NVIDIA Spark and an NVIDIA Jetson Orin Nano. Everything running locally, without relying on third-party paid services. Nothing is more secure than keeping everything at home.

Homelab & Docker Local Generative AI Edge AI n8n & RAG NVIDIA Jetson / Spark Cybersecurity
  View full report (121 pages)   View architecture
Project developed together with Óscar Toledo Fernández, ASIX classmate and co-author of the technical report. The homelab, n8n integration and local AI (PC) are Izan's main responsibility; the Raspberry Pi + AI HAT, the NVIDIA Spark and the Jetson Orin Nano were developed jointly.
5
Hardware platforms
7
Services in production 24/7
121
Technical report pages
0€
Cost of paid cloud services

Executive summary

Vision

Novacreux was created to bridge the gap between cutting-edge AI and robust security and data sovereignty. The premise is simple: a company or professional does not need to hand over their data to third parties to benefit from generative AI, automation or advanced monitoring — they can have everything on their own network. We believe that a secret leaving a secure place can create a security breach, which is why secrets never leave home.

Technical objective

Design, deploy and document the same AI and automation ecosystem across very different hardware tiers — from a low-power Raspberry Pi to a NVIDIA Spark de nivel profesional — para entender qué plataforma conviene a cada caso de real-world use case: cost, performance, energy consumption and scalability. This project also demonstrates that, as currently built, it can be easily scaled.

Ecosystem architecture

The project is divided into five independent blocks that share the same philosophy (local-first, with no external dependencies) but address different needs: core infrastructure, desktop AI, low-power embedded AI and professional-grade AI.

NOVACREUX Homelab Docker · Ubuntu Server Local AI (PC) RTX 4060 Ti 16GB Raspberry Pi 5 + AI HAT+2 (Edge) NVIDIA Spark Blackwell · up to 128GB Jetson Orin Nano Robotics · IoT n8n — Agente IA con RAG izan1706.github.io (este portfolio)
Five independent hardware blocks, united by the same philosophy: local AI, with no external dependencies. The n8n RAG agent connects directly to this same portfolio.
Block 1 · Core infrastructure

Homelab & Containerization

Behind any local AI service, infrastructure is needed to support it. A reused PC (Dell OptiPlex, Intel i7-4790, 16GB RAM, 436GB SSD) runs Ubuntu Server without a graphical interface, with each service isolated in its own Docker container: if one fails, the rest keeps running.

Servidor Dell OptiPlex usado como homelab
The homelab server itself: reused hardware, 6W average power consumption, running 24/7.
PC IA Local Other devices Other devices Switch Home Router Homelab
Physical network topology of the homelab (recreated from the technical report).

Deployed services — click each one to see the details:

VPN · Tailscale

Secure remote access to the entire network without exposing ports, with access control and audit logs.

Pi-Hole

DNS propio de la red doméstica: bloqueo de publicidad y dominios maliciosos, nombres locales fáciles de recordar.

Immich Gallery

Self-hosted alternative to Google Photos, with mobile synchronization and local machine learning.

Fail2ban

Brute-force intrusion prevention: detects failed login patterns and automatically bans the IP.

Uptime Kuma

Visual monitoring of services, TCP ports and containers, with alerts for incidents.

Netdata

Real-time visibility into CPU, memory, disk and network — identifies bottlenecks instantly.

n8n

Low-code automation orchestrator, a free self-hosted alternative to Zapier.

Block 2 · Desktop AI

Local Generative AI (PC)

The goal is to demonstrate that, with consumer hardware, it is possible to have professional-grade AI workflows with minimal latency and full control over the models — without any data leaving the local network. The stack runs on Ollama as the inference engine and Open WebUI as the interface, all in Docker containers.

DeviceIzan's Personal PC
CPUIntel i5-13600KF
RAM2 × 8GB Kingston Fury 3200MHz
GPUZotac NVIDIA RTX 4060 Ti — 16GB VRAM
Disco1TB NVMe M.2 Kingston
Performance12–18 tokens/s con modelos de hasta 24B
  • Inference engineOllama
  • InterfaceOpen WebUI (Docker)
  • Critical requirementVRAM ≥ 16GB
  • SOLinux (CUDA)
  • IsolationDedicated users and permissions
Block 3 · Low-power Edge AI

Raspberry Pi 5 + AI HAT+2 — Video Analyzer

An 8GB Raspberry Pi 5 with the AI HAT+2 module (Hailo-8 NPU, 26 TOPS) runs a pipeline in Python with OpenCV and the Hailo API capable of describing video content: in a real-world environment this translates into detecting workplace hazards, smoke or objects without human intervention, and automatically triggering protocols or notifications.

Raspberry Pi 5 con módulo AI HAT+2
Raspberry Pi 5 + AI HAT+2: 26 TOPS de NPU en un consumo de 5-12W.
  • BoardRaspberry Pi 5 — 8GB RAM
  • AcceleratorAI HAT+2 (NPU Hailo-8, 26 TOPS)
  • Average power consumption5–12 W
  • AI performance3–5 tokens/s (modelos de 3B)
  • Software stackPython venv · OpenCV · hailo-python-api · Ollama
  • Use caseComputer vision / Edge IoT
Blocks 4 & 5 · Professional AI and robotics

NVIDIA Spark & Jetson Orin Nano

At the opposite end of the spectrum from the Raspberry Pi are NVIDIA's two platforms: the Spark, designed for large-scale training and inference, and the Jetson Orin Nano, focused on Edge AI and robotics with agents that act on sensors in real time.

NVIDIA Spark

Blackwell architecture in a personal supercomputer form factor, tested through remote access to a virtual machine.

  • GPUBlackwell
  • Memory48–128 GB
  • StorageNVMe hasta 60TB
  • Power consumption100–240 W
  • AI capacityModelos de 70B a 400B
  • Performance15–22 tokens/s
Kit de desarrollo NVIDIA Jetson Orin Nano
Jetson Orin Nano development kit: sensors, actuators and touchscreen for prototyping AI agents on physical hardware.

NVIDIA Jetson Orin Nano

Using a development kit with sensors, actuators and an integrated display, we tested how an AI agent can read sensor data and act accordingly and autonomously — the foundation of any robotics or industrial IoT application.

  • ArchitectureAmpere
  • Performance40 TOPS
  • Power consumption7–15 W
  • Shared memory8 GB
  • AI capacityModelos de 1B a 7B
  • AI performance10–15 tokens/s

Hardware comparison

Four platforms, four different purposes. None excels at everything — each has its place depending on the use case.

RTX 4060 Ti 16GB Raspberry Pi + HAT2+ NVIDIA Spark Jetson Orin Nano
Architecture Ada Lovelace ARM Cortex-A76 + NPU Hailo-8 Blackwell Ampere
Memory 16GB GDDR6 8GB LPDDR4 + 26 TOPS NPU 48–128GB 8GB compartida
Power consumption (TDP) 165W 5–12W 100–240W 7–15W
Typical use IA local (hasta 24B), gaming Edge AI, visión artificial, IoT Entrenamiento, modelos de 70B–400B Edge AI, robótica
Advantage Low cost, low latency Ultra-low power consumption, portable Scalability, large memory JetPack ecosystem, real-time
Limitation Insufficient VRAM for large models Not suitable for large models Cost, power consumption, complexity Limited shared memory (8GB)

  Energy cost calculator

Estimated annual cost per platform, based on the power consumption measured in the project.

—
Daily consumption
—
Annual consumption
—
Estimated annual cost

Lessons learned and next steps

What worked

  • Segmenting each service in Docker prevents a failure from bringing down the rest of the homelab.
  • Comparing 5 real platforms teaches more than any specification table: each one has a clear use case.
  • Using the project itself (the n8n agent) in production on this portfolio, not just as a demo.
  • Measure actual power consumption and cost before recommending a platform, not just performance.

Next steps

  • Live Raspberry video analyzer (real-time camera, not pre-recorded video).
  • Migrate the VPN from Tailscale to 100% self-hosted WireGuard.
  • Replace Fail2ban with CrowdSec, with shared intelligence between nodes.
  • Raspberry chatbot accessible via WhatsApp/Telegram, orchestrated with n8n.
  • Set up a NAS with real redundancy for the homelab.

Want to see the project in more detail?

The complete technical report covers every installation, configuration and testing step for the 5 project blocks.

  Download report (PDF)   Back to projects

Systems Administrator and Cybersecurity Specialist focused on robust, secure and automated solutions.

Contact

  • izanro06@gmail.com
  • Barcelona, España

Quick Links

Systems Operational Uptime: 99.9%
© 2026 IZAN. All rights reserved. Technical portfolio for systems administration and cybersecurity.
SSL Secured | AES-256 Verified Connection