From network technician to AI architect, without skipping a step
I'm Mircea Gabriel Pincovai, a Cloud, Platform & AI Architect based in Manresa (Barcelona). Over 10 years designing enterprise architectures and large-scale cloud platforms, and over 8 years designing AI architectures and solutions — from Machine Learning, NLP and Computer Vision to today's Generative AI, LLMs, RAG, AI Agents and Model Context Protocol (MCP) architectures.
Today I combine three roles: Founder & CEO and Principal AI Architect at Nexunia, co-founder of Avvanex, and Cloud Tech Lead at adm Group Ltd. All three share the same obsession: that a system does exactly what it says it does, in an auditable way, and holds up under real load, not just in the demo.
A native Spanish, Catalan and Romanian speaker with professional working English, I started repairing computers and configuring networks in 2015, and I haven't stopped climbing the abstraction ladder since: from systems technician to hybrid infrastructure administrator, from there to cloud architect, and from there to designing the AI platforms that run agents in production today.
Eleven years, one standard
January 2026 — Present
Founder & CEO, Principal AI Architect
Nexunia
Technology vision, architecture and evolution of a multi-tenant AI SaaS platform: agents orchestrated with models, tools, APIs and enterprise systems via Model Context Protocol (MCP) and RAG, with tenant isolation, observability and high availability built in.
Ongoing, in parallel
Co-founder
Avvanex
Digital growth agency for service businesses. Nexunia was born inside Avvanex as its first case study before spinning out as its own product.
March 2021 — Present (5 years 8 months)
Cloud Tech Lead
adm Group Ltd
Cloud architecture standards and governance models for the organization, working alongside Engineering and DevOps teams. Previously Cloud Architect at the same company (2021-2022).
June 2020 — May 2021
Cloud Architect
Revertis
IaaS/PaaS/SaaS on Microsoft Azure, Office 365 and security/compliance, VMware and Hyper-V virtualization, SQL Server, networking, and on-premise-to-cloud migrations.
July 2019 — June 2020
Cloud & Security Consultant
CleverTask IT Solutions
Office 365, Exchange, SharePoint and Teams administration, PowerShell automation, on-premise/cloud project management and migrations, data governance, and 2nd/3rd-level incident resolution.
August 2018 — July 2019
Cloud & Hybrid Infrastructure Administrator
Kids&Us
Microsoft Azure, Office 365 and Google Admin, Hyper-V virtualization, NAS storage, network monitoring (Nagios, PRTG), and level 2/3 helpdesk.
October 2015 — September 2018
Systems Lead / Systems Administrator
TecnoConverting Engineering
Wintel and Linux infrastructure, VMware virtualization, Veeam backups, networking, and early forays into web development (HTML5, CSS3, JS, PHP).
March 2015 — August 2015
IT Technician
Confimatik
Network administration, service installation and configuration, equipment maintenance and repair — where everything above started.
The formal part, for whoever needs to verify it
Certifications
- AZ-303: Microsoft Azure Architect Technologies
- Certified Kubernetes Administrator (CKA)
- CCNA Discovery v4
- Windows Server 2016: System Virtualization & High Availability
Education
Master's in Cloud Computing: Azure, AWS & GCP
CICE, La Escuela Profesional de Nuevas Tecnologías · 2019 — 2020
Higher Technician in Network Systems Administration
IES Lacetania · 2016 — 2018
Vocational Diploma in IT
IES Lacetania · 2013 — 2015
Languages
- SpanishNative
- CatalanNative
- RomanianNative
- EnglishProfessional
- HungarianElementary
How I decide, not just what I build
Auditable by design
Everything an AI agent executes gets logged: what was called, with what data, with what result. If it can't be audited, it isn't trusted.
Staging before production, no exceptions
No change reaches production without going through staging and being validated there first — a rule I apply since seeing the real cost of skipping it firsthand.
Strict isolation between customers
In a multi-tenant system, customer identity is always resolved from the authenticated session, never from data the customer themselves could manipulate.
AI interprets, the system decides
A language model never executes a business action directly — it interprets intent and requests a tool; the backend validates, executes, and confirms.
I build things that hold up
I'm not interested in the prototype that impresses in a demo and falls over with the first real user. I'm interested in the system that's still running — and that you can still explain — a year later.