Industrial EngineerUPC · ETSEIBHead of AI at BTECH
Industrial engineering and artificial intelligence. In practice, almost everything I do starts with someone describing a symptom.
We’re at the start of a new industrial revolution — the AI one. My place in it is clear: to integrate it into everything that already exists and works, and to open up processes that weren’t possible before —digesting and cross-referencing far more information than fits in one head—. Used well, it makes us ultra-capable; used badly, it takes away more than it gives. The difference is judgement. Build the future, don’t wait for it.
Engineers at the company where I built the AI department
3 months
Of one person’s annual work, cut down to hours on a single process
CEO
Direct reporting line to the CEO
GB10
Own compute, on-premises and on-edge: data never leaves the company.
At a glance
Sovereign AI, secure by design
Models that run on-premises: not a single datum leaves the building. On RKE2, the DISA STIG–certified Kubernetes for defence and intelligence (FIPS 140-2), plus agentic systems demoed live.
NVIDIA stack, end to end
From CUDA and GB10 to NeMo, NIM and Dynamo. Distributed inference on Kubernetes, a contribution to OpenShell, and part of the Hugging Face open-source community.
Physical AI and simulation
Digital twins, finite elements and models that reason over physical systems. From calculation to the real world —down to open-source robotics (Pollen Robotics · Hugging Face) I build for fun.
AI for biology and pharma
Computational biology with AlphaFold (up to its latest versions), a glucose-energy thesis for implantable devices with a patent process started, and experience in a regulated pharmaceutical environment.
Toward a brain digital twin
Where all of the above is heading: digital twins and physics-informed AI on GPU to simulate biology instead of assaying it, and test drugs in silico. Rooted in real bioengineering —the citric-acid cycle and glucose energy for implantable devices— and, from 2026, in the study of the brain: how it learns, decides and fails.
Greatness is not intelligence. Greatness comes from character. And character isn’t formed out of smart people. It’s formed out of people who suffered.
Jensen Huang, NVIDIA · Caltech, 2024
Above all, I consider myself a lifelong learner: I learn mostly from the people around me, I keep curiosity as a working method, and I am well aware that there is still a long road ahead.
What brings you here?
Then what matters to you is the method, not my CV. Scroll down and see, case by case, where I start.
Pick one and I take you straight to what you are after. In the end, my job is solving problems; yours included.
Diagnosis
It almost always starts with a sentence like this.
I have heard every one of them more than once, in boardrooms and on the plant floor. None describes the problem: they describe the symptom.
We have data everywhere and decisions still get made on instinct.
What it usually is
The data is not missing: it is unassembled. Each area measures what it owns by its own criteria, so in the meeting where something has to be decided there is no single figure, there are five versions of it and none can be defended.
How it gets tackled
Before automating anything, agree what gets measured and who answers for each number. Then one place where the metric computes itself and can be audited. The system does not decide: it assembles the metric so that whoever has to decide, can.
We have the plan, the budget and the technology. And the project is not moving.
What it usually is
The plan almost never fails. What fails is that nobody has explained to the people who must run it why it is in their interest, so the project competes against the usual work and loses every week.
How it gets tackled
Start with the area that gains most from the change and the process that hurts most, not the one that is easiest to touch. One case that works convinces more than any presentation, and from there the rest ask to join. The hard part is not the architecture: it is getting the organisation to want to use it.
To test a change we have to stop everything, so we do not test it.
What it usually is
The cost of the test is not the change: it is the shutdown. So decisions get made without trialling, and what reaches production is the first version nobody was able to question.
How it gets tackled
Stand up a replica of the system where the change can be trialled without touching the original: simulated physics when a machine sits on the other side, a replica of the process when it is an organisation, and in both cases models trained on the system’s own data. The chain is the same whether it is a production line, a technical building, a living system or the way an entire company works.
When someone leaves, half the documentation walks out in their head.
What it usually is
It is not unwritten out of neglect: it is written in fifteen places and nobody knows which version is current. Searching costs more than asking, so people ask, and so the knowledge goes on living in people.
How it gets tackled
One place with a current version and permissions, and on top of it a layer that answers by citing the document it drew the answer from. The moment searching costs less than asking, people search on their own.
We want to use AI, but not one piece of data leaves this building.
What it usually is
That is why most trials stay trials: the pilot gets built outside, it works, and then there is no way to bring it back in.
How it gets tackled
Build it inside from day one: open models, in-house infrastructure and in-house GPU compute. A pilot born outside that then has to be repatriated almost never is, because by then it depends on things that cannot be moved.
We want to move forward with AI, but nobody can tell us what the law lets us do.
What it usually is
The European regulation barely prohibits: it classifies. What paralyses is not the rule, it is that nobody has said which risk tier each use case falls into, so out of caution they all get treated as the worst and none gets built.
How it gets tackled
Classify case by case before building anything: unacceptable, high, limited or minimal risk, which is what actually decides the obligations. Most internal uses fall in the two lower tiers, and knowing that unblocks the project without lowering the guard on the one that genuinely is high risk.
We can’t use a cloud AI with our data.
What it usually is
In regulated sectors —defence, health, banking— the problem isn’t the AI, it’s the data’s chain of custody: every call to an external service is a copy leaving, and that fails an audit or a tender.
How it gets tackled
Everything on-premises and auditable: open models on hardened Kubernetes (RKE2, FIPS 140-2/STIG), your own GPU and agents secure by design, with a trace of every action. The AI comes in without the data going out.
The distance between the first column and the second is where the budget goes, and most of my work is closing it before touching anything. What I deliver in the end is not a tool: it is a figure you can decide on, produced in hours instead of by hand over months.
Experience with
Barcelona Technical CenterFI GroupFnacGrifolsFòrum ETSEIB
Areas of expertise
Seven fronts, the same method.
The disciplines change; where you start does not. From structural analysis to the architecture of artificial intelligence systems, the order is always the same: understand, measure, and only then touch.
01
Consulting
Process
Working inside someone else’s problem: understanding it, documenting it and defending it before whoever has to approve it. At FI Group, face to face with the client and with no intermediaries.
Public grantsTax deductionsClient management
02
R&D and innovation
Process
Almost every industrial company does development the law recognises as R&D, and almost none claims it: what is missing is the report that proves it. That report is engineering work, not paperwork — you have to understand what was built to argue why it qualifies.
Research projectsInnovation
03
Artificial intelligence
Digital
A department built from nothing at a 400-engineer company: RAG architectures, autonomous agents and model fine-tuning on in-house infrastructure and GPU compute, with no data leaving the company.
RAGOn-premise LLMsData architecture
04
Finite elements
Digital
The full analysis chain on automotive components: ANSA for pre-processing, Pam-Crash for dynamics, Abaqus for static analysis and Meta for post-processing.
ANSAPam-CrashAbaqusMeta
05
Energy
Physical
A master’s in Energy Engineering and a postgraduate in data centre energy management: renewables, efficiency, smart grids, storage and hydrogen.
PhotovoltaicsSmart gridsStorageHydrogen
06
Industrial
Physical
Plant projects end to end: an 86 kW self-consumption solar plant, an exhaust gas system and ERP architecture.
PlantERPISO 50001
07
Systems and IT
Digital
Access and network management, ERP roll-out, biometric control, encryption and automation of internal processes.
NetworksBiometricsEncryptionAutomation
Method
From the real system to the decision.
The thread running through everything I do: turning something that already works into a model you can trust, so scenarios can be tested without touching it. It makes no difference whether that something is a plant or an entire company.
01 · Real
Real system
A plant, an installation, a supply chain or the way an entire company works. The starting point is always something that exists and behaves.
02 · Measurement
Data
Sensors, load curves, historical series or the records the systems already generate. Whatever says how it really behaves, not how it should.
03 · Model
Digital twin
An analysis mesh, a trained model or a replica of the process: whatever reproduces that behaviour within a bounded, measured error.
04 · Use
Decision
Simulate scenarios and decide on the model, not on the system that is running and that somebody depends on.
From analysis engineer to building an AI department.
I am Guillem Navarro Vargas, an Industrial Technologies engineer from BarcelonaTech (UPC-ETSEIB). I started out meshing automotive components and today I lead the artificial intelligence department of a 400-engineer company.
That path explains how I work: understand the whole system before optimising any of its parts. I learned it in analysis, where one badly placed assumption invalidates the result however fine the mesh is.
And I didn’t get here alone, nor do I have it all figured out: I learn fast, ask without ego, and surround myself with people who know more than I do in their field. More often than not, the credit belongs to the team.
How I work
The seven pillars of the journey.
Every company along the way left me one pillar, and I went after each on purpose. They rest on the psychology of problem-solving and on the soft skills you can’t compute, and they all serve the same strategy: understand the whole system —from the physical world to the model— before touching a single part.
兵法：一曰度，二曰量，三曰數，四曰稱，五曰勝。地生度，度生量，量生數，數生稱，稱生勝。
The art of war has five steps: measurement, estimation, calculation, balance and victory. The ground gives rise to measurement; measurement to estimation; estimation to calculation; calculation to balance; and balance to victory.
Sun Tzu · The Art of War, ch. IV (5th c. BC)
Twenty-five centuries on, it’s still the same order: measure, estimate, calculate, decide. My pillars just give it the name today’s psychology uses.
01Problem representation
The whole system, before the parts
I won’t touch a part until I understand the system holding it up. Experts read a problem by its structure, not its surface (Chi, Feltovich & Glaser, 1981); I trained that eye running an industrial plant end to end — networks, ERP, IoT — where nothing works in isolation.
02Analogical transfer · Physical AI
From the physical world to the model
I solve a new problem by carrying the solution over from another domain: that is analogical transfer (Gick & Holyoak, 1983), the engine of expert creativity. From finite-element analysis and the electric vehicles I repaired with my hands at Robotic Dreams to the models that now reason over physical systems: that, to me, is physical AI.
03Metacognition
The assumption decides the result
I police my assumptions as hard as I run the numbers. Expertise is as much metacognition as technique (Flavell, 1979): in finite elements a bad assumption invalidates the result no matter how fine the mesh — and that holds for a mesh or a business decision alike.
04Deliberate practice
Deliberate practice, not hours
I choose where to work for what I’ll learn, not for what I already know. Expertise comes from practising what you haven’t mastered (Ericsson, 1993); so every company and degree was a decision, and I keep to the frontier: first cohort of Perplexity’s AI Business Fellowship.
05Communication · soft skill
Turning the technical into a decision
A result you can’t defend doesn’t exist. The expert forgets what not-knowing feels like — the curse of knowledge (Camerer et al., 1989) — so I translate on purpose: I defended R&D dossiers before clients and the tax authority at FI Group, and demoed agentic systems live to a non-technical congress at DETCON.
06Leadership · soft skill
With and through people
I lead by shared mental model, not by org chart. The best teams think in the same frame (Cannon-Bowers & Salas, 1993); I’ve built that leading communications for Fòrum ETSEIB, broadcasting live for Ràdio Calella Televisió, lending a hand at the European Blockchain Convention, and building an AI department from scratch.
07Bounded rationality
Rigour under constraint
I decide with rigour when time and resources are finite — which is always. That is the real definition of rationality (Simon, 1955), and I sharpened it at Grifols, managing access in a regulated pharmaceutical environment where error is not an option.
And I know what I don’t know: every project reminds me how much is left to learn — and that what turns out well is almost never down to one person alone.
Career and companiesEducation
03/2016 — present Ongoing
Industrial company · IT engineering and automation
Networks, IoT, ERP and the plant’s digital backbone. Held alongside everything else ever since.
09/2017 — 09/2018
Robotic Dreams S.L. · Electric vehicle repair
Pineda de Mar. Diagnosing and repairing electric vehicles: batteries, motors and power electronics learnt hands-on.
From research to the global market: technology transfer, funding and deep-tech scaling.
In progress
What I am leading right now.
Corporate ecosystem with agents
BTECH · Project lead
It is not one more tool bolted onto the ones already there: it is the layer where all of them meet. Commercial management, resource planning, knowledge base, training, engineering and simulation live in one ecosystem, talk to each other over declared paths and share a single identity. On top, a layer of agents with their own tools that can query any of them, cross-check what one says against another, and act on them where it is entitled to. It reaches every department without exception: leadership, finance, human resources, sales, quality, legal, engineering and production.
Everything under one roof
A company’s systems usually speak different languages and none of them knows what the others are doing; when someone needs to cross two, they do it by hand. Here each program communicates with the rest — and with whatever external ones are needed — over a declared path, not through loose integrations nobody remembers who built. That is what turns a collection of applications into a system.
Auditable infrastructure
Everything runs on RKE2, the Kubernetes distribution validated to FIPS 140-2 and compliant with DISA STIGs: the one admitted into US federal environments. It is not security decoration. It is what lets an outsider come in and check what encrypts what, who accessed what and when, without the answer being “we would have to look into that”.
The agent sees no more than you
It is the same agent for the whole workforce, but it inherits the permissions of whoever invokes it: when a person uses it, it reaches exactly what that person reaches, not one record more. It looks like an implementation detail and it is the piece that decides whether the system can exist at all: without it, a corporate assistant is a data leak with a good interface, and one well-phrased question is enough to pull up somebody’s salary.
Local or external, case by case
A router decides, query by query, whether it is resolved by a model running in-house or one outside. Sensitive material never leaves; what is not sensitive can use the best model available at that moment. Data sovereignty stops being a position defended in meetings and becomes a rule that enforces itself.
A world to test in
Inside the same ecosystem lives a simulation environment where engineering geometry, physics and the models that learn from it share a single scene. Humanoid robotics is trained there against a simulated world before the robot exists: situations are generated that would take months of data collection on a real plant floor, and models that understand image and language at once learn to read them. It is a test bench for something that has not been built yet.
Every flow in view
Everything moving from one program to another is visible, measurable and can be cut. A capability is opened to one department and closed to another; a tool is granted to an agent and taken back; a request is followed from whoever made it to the last system it touched. Without that traceability there is no governance, only trust, and trust cannot be audited.
A system like this is, in practice, a digital twin of the company. Not of a part or a plant: of how information moves, who can see what, and how decisions get made. And a twin is good for the same thing here as in engineering — trying the change before making it, and knowing what breaks if it goes wrong.
There is usage data, but it is not public. Neither is the ERP, the commercial system, the simulation environment or the specific architecture.
Tools
What I work with.
What I use to deliver, grouped by domain and ordered by fluency. Not everything I have touched: everything I can stand behind.
NVIDIA DGX Spark (GB10) · Kubernetes · RKE2 · CUDA · Argo CD · GitLab · Coder · Docker · Istio · Python · Linux and Linux Server · Windows Server · Git · LaTeX · NVLink · NCCL · Dell servers
Systems and platforms
Active Directory · authentik · SAP · Azure · Citrix · VPN · Access control · n8n · Wiki.js · Drupal · Jira · Power Automate
Energy and plant
ISO 50001 · Self-consumption solar · Industrial IoT · ERP · Minitab
Management & people
Team leadership · Project management · Hiring & onboarding · Executive communication · Negotiation · Mentoring · OKRs & KPIs · Cross-functional work
Trademarks, logos and product names mentioned belong to their respective owners. They are cited solely to describe professional experience with them and imply no commercial relationship, sponsorship or endorsement by their owners.
Credentials
Where I trained.
Degrees, certifications and what is under way. What points where I am heading is the postgraduate with Schneider Electric and the NVIDIA certifications.
Education
Postgraduate in Data Centre Energy Management · Schneider trackUPC School · Schneider Electric
Master’s in Energy EngineeringUPC · ETSEIB09/2022 — 05/2025
Master’s in Industrial EngineeringUPC · ETSEIB09/2022 — 05/2025
Bachelor’s in Industrial Technologies EngineeringUPC · ETSEIB09/2018 — 07/2022
Large Language Model Agents (MOOC)UC Berkeley01/2025 — present
Course in Deep Tech to Market Not started yetUPC School · Fractus-UPC Deep Tech HubFrom 11/2026 · 1 ECTSFrom research to the global market: turning disruptive science into high-impact companies. Technology transfer, funding and deep-tech scaling.
Psychology degree Not started yetUNED · Spain’s national distance universityFrom 10/2026The neural networks I use every day are an abstraction of something that actually exists. Understanding the original — how a brain learns, how it decides and above all how it fails — is not a detour from what I do: it is the other half of the same problem. And there is a second reason, learned from leading: a project’s bottleneck is almost never technical, it is that there are people who have to be understood before they can be led.
Certifications
NVIDIA-Certified Associate · Generative AI LLMsNVIDIA2025
Deploying RAG Pipelines for Production at ScaleNVIDIA Deep Learning Institute2025
Mobile application developmentUniversidad Complutense de Madrid2020
Drone pilotAESA2023
Talks
Agentic systems applied to private investigation
DETCON 2026 · Private investigators congress, Barcelona · 2026
Invited speaker as an artificial intelligence specialist. Agents running locally and in isolation, from an ordinary laptop to an NVIDIA Jetson Orin, with no data leaving the machine. Live demonstration of the whole flow: input over messaging, automatic drafting of the report and follow-up questions on it. Presented as beta-stage technology rather than production, and as the direction the market is taking.
Technical community
AutoCFD · Automotive CFD Prediction Workshop
International working group · Monthly meetings · 5th edition
I take part in the monthly meetings of the AutoCFD working group — the international workshop that measures how well CFD actually predicts on automotive geometries. My focus there is the road ahead: simulation through machine-learning models, trained on open, high-fidelity CFD data, that predict in seconds what takes a solver hours. It is exactly where classical simulation and physical AI meet —the same crossing I work at every day.
Perplexity · AI Business Fellowship
Program community · First cohort
Part of the technical community of Perplexity’s AI Business Fellowship: a small network focused on bringing AI to real business decisions.
Hugging Face
Open-source community · Models, datasets and robotics
Part of the Hugging Face open-source community: open models and datasets, and the Pollen Robotics hardware I tinker with.
Licences Driving licences A1, A and B · Recreational boat skipper · Drone pilot (AESA) · Open Water Diver (SSI, up to 18 m)
Résumé
The track record, on one page.
A summary of the whole track record, condensed into a single page ready to print or take away as a PDF.
Industrial Technologies Engineer from UPC-ETSEIB, with master’s degrees in Energy Engineering and in Industrial Engineering and a postgraduate in data centre energy management from UPC School with Schneider Electric. Head of AI at BTECH, where I built the AI department of a 400-engineer company, reporting directly to the CEO. I work where critical infrastructure, energy and artificial intelligence applied to physical systems meet, with one fixed principle: the machine assembles the metric, the decision stays with a person.
Experience
Head of AI
Built the AI department of a 400-engineer company: I proposed the business case, defended it to the leadership and set the area up from nothing, reporting directly to the CEO. Team leadership and cross-departmental work with the rest of the organisation, starting with finance: process automation and KPI and indicator systems, with cases where three months of a person’s annual work came down to hours. A significant part in bringing in Regulation (EU) 2024/1689 and the management framework aligned to ISO/IEC 42001, along with the internal training on both. An AI ecosystem integrated into simulation and CAD workflows, with LoRA-fine-tuned LLMs, RAG, autonomous agents and memory systems on Kubernetes and in-house GPU compute.
CAE Engineer and AI Integration Lead
Finite element analysis on automotive components with ANSA, Pam-Crash, Abaqus and Meta, and integration of LLM engines and RAG architectures into the simulation workflow.
IT Engineer and Process Automation Specialist
Held alongside the rest of my career. Optimisation of distribution, data transfer, backups, updates and IoT integration. Key part of developing an 86 kW photovoltaic plant and of ERP roll-out across industrial and logistics operations.
R&D Consultant
Technical reports and direct client relationships. Coordination of grants, tax deductions and social-security rebates, both self-assessed and certified.
IT Access Control Management · internship
Joined as a level-2 helpdesk technician and was promoted to access control management: systems administration, Active Directory and SAP.
Corporate Communications Director
Led the corporate communications area and the team within it: negotiation and company relations for the school’s careers forum. The first team I led.
Education
Postgraduate in Data Centre Energy Management · Schneider track
Critical infrastructure end to end: advanced energy management of data centres, innovation in cooling and control systems, and the design and operation of hybrid AC/DC microgrids and uninterruptible power systems.
Master’s in Energy Engineering
Master’s in Industrial Engineering
Bachelor’s in Industrial Technologies Engineering
Large Language Model Agents (MOOC)
Course in Deep Tech to Market
Psychology degree
Certifications
NVIDIA-Certified Associate · Generative AI LLMs
Deploying RAG Pipelines for Production at Scale
Certified SolidWorks Associate
Talks
Agentic systems applied to private investigation
Community
AutoCFD · Automotive CFD Prediction Workshop
Perplexity · AI Business Fellowship
Hugging Face · open-source community
Competencies
ConsultingProcess
R&D and innovationProcess
Artificial intelligenceDigital
Finite elementsDigital
EnergyPhysical
IndustrialPhysical
Systems and ITDigital
Other details
Software
Python · ANSA · Abaqus · Pam-Crash · Meta · SolidWorks · MATLAB · Kubernetes · Linux
Licences
A1, A and B · Recreational boat skipper · Drone pilot · Open Water Diver (SSI, up to 18 m)
Languages
Catalan and Spanish native · English C1 certified · French A2 certified
Contact
Tell me the problem. I answer with the method.
Feasibility, approach and an estimate of hours. Within 48 working hours, written by me and with no commitment.
