Skip to content
Hugo Romero

Hugo Romero · AI Engineer & Architect

The Forward Pass

My background, drawn as a neural network. Each circle is a stage of my career; the lines show what carried over.

Network diagram of Hugo Romero's career. Read the Layers section for the full text version.

Each neuron is a role or a foundation; use Tab to move through them in layer order and Enter to open one, or read the Layers list.

Fig. 1, explained

How to read it

No AI background needed. A neural network learns by passing a signal through layers of simple units, each one building on the last. That turned out to be a good way to describe a career. Here is what each part means, and what it means for me.

  1. Inputs

    input features

    A network starts from raw signals, and everything it learns is built on top of them.

    Mine: math olympiads, a Computer Science and Business degree, coding contests, volleyball and music.

  2. Connections

    weights

    Lines carry signal from one circle to the next. The thicker the line, the more of a skill carried over.

    The thickest: Terraform learned as an intern at MásMóvil became the backbone of Orange's AI platform.

  3. Layers in between

    hidden layers

    Each layer takes what the previous one produced and turns it into something more useful.

    Mine went from research to infrastructure to AI engineering to architecture, each built on the last.

  4. Lighting up

    activation

    A circle lights up only when enough signal reaches it. Below the threshold, it stays grey.

    At Orange this one fired hard: ML pipelines, LLM governance and AI agents, all running in production.

  5. The answer

    output · softmax

    The last layer turns everything before it into one prediction, with a confidence score.

    Today that's Founding Engineer at Feniria Labs. The network's verdict: AI Engineer & Architect, 97% sure.

  6. Key

    • FoundationsResearchInfrastructureAIProduct / Business
    • Dashed circle: tried, then let go dropout
    • Small square: a trait that nudges a whole layer bias
    • Arc: a skill that skipped ahead residual
    • Gold dots: signal moving forward Play
    • Pink dots: lessons flowing back Lessons

Fig. 1, as text

Layers

The same story as a plain list, from today back to the beginning. Open any role for the details, the lesson that stage taught me, and what it carried forward.

  1. Layer 06 · 2026 – present

    Output

    Founding engineer: inference optimisation, product and the first paying customer.

    Lesson · Customers decide what the product is

    Trait · Chess-player

    Founding Engineer · Feniria Labs

    Jun 2026 – present

    Product, LLM inference and commercial execution at an early-stage GPU optimisation company.

    • Built automated vLLM optimisation: scheduling, batching, concurrency, KV cache and GPU allocation
    • Built benchmarking for TTFT, throughput, latency and GPU utilisation
    • Led the pilot-to-contract conversion with Feniria's first paying customer
    • Built an ICP engine and CRM that fed 400+ qualified prospects into sales
    • Python
    • vLLM
    • GPU
    • AI Agents
    • React
    • Vite

  2. Layer 05 · 2026

    Architecture

    Setting company-wide AI standards and the agent architecture behind them.

    Lesson · Standards scale further than heroics

    Trait · Stoic principles

    AI Architecture & Innovation Tech Lead · Orange

    Mar 2026 – Jun 2026

    Led a transversal team turning business needs into AI architecture, company-wide.

    • Defined AI architecture standards and best practices across engineering
    • Deployed production agent systems on A2A and MCP
    • Integrated Gemini Enterprise, Slack and Microsoft Teams as agent surfaces
    • Reduced AI OPEX through model routing, caching and governance
    • A2A
    • MCP
    • Gemini Enterprise
    • ADK
    • LangGraph
    • GCP
    • ArgoCD

    Carried forward

    • Feniria Labs: Cutting AI OPEX at Orange is the problem Feniria sells a fix for: cheaper inference.

  3. Layer 04 · 2023 – 2026

    AI Engineering

    MLOps, LLMOps and agentic systems for a large enterprise, in production.

    Lesson · Cost is a design constraint

    Trait · Curiosity about the latest AI

    AI Engineer · Orange

    Sep 2023 – Mar 2026

    Production GenAI and agentic systems, plus the platform underneath them.

    • Built MLOps pipelines on GCP covering training to monitoring
    • Wrote and maintained the Terraform modules behind the AI infrastructure
    • Implemented LLMOps governance: prompt versioning, budget control and RBAC
    • Built agentic workflows with MCP servers and multi-provider LLM access (Azure, GCP, OpenAI, LiteLLM)
    • Python
    • GCP
    • Terraform
    • Kubernetes
    • LangGraph
    • MCP
    • LiteLLM
    • Vertex AI

    Carried forward

    • Orange: Building the platform is what earned the mandate to set its standards.

  4. Layer 03 · 2022 – 2023

    Infrastructure

    Cloud-native engineering at telecom scale, from Terraform to reactive Java in production.

    Lesson · Ship daily, merge daily

    Trait · Competitive sport discipline

    DevOps & Data Intern · MásMóvil

    Oct 2022 – Sep 2023

    Infrastructure and automation for a cloud-native data team with a 5 PB platform.

    • Deployed GCP resources with Terraform and CI/CD pipelines
    • Automated ETL workflows with Apache Airflow
    • Automated the data team's daily operational processes
    • Terraform
    • Airflow
    • GCP
    • Kubernetes
    • CI/CD

    Carried forward

    • Orange: Terraform learned as an intern became the modules behind Orange's AI platform.

    Backend Developer · MásMóvil

    Mar 2023 – Sep 2023

    Reactive Java on a cloud-native platform shared by 500+ developers.

    • Merged to production daily on the MasStack microservices platform
    • Built reactive Java workflows with Cadence
    • Integrated with government SOAP APIs in a high-volume, high-risk domain
    • Reactive Java
    • Cadence
    • Kubernetes
    • Bazel
    • Microservices

    Carried forward

    • Orange: Daily production merges set the bar: ship AI, do not just demo it.

  5. Layer 02 · 2019 – 2023

    Research

    First jobs close to the university: product discovery, game theory and evolutionary algorithms.

    Lesson · Build it from scratch once

    Trait · Critical thinking

    Digital Business Developer · Grupo SM

    Jun 2019 – Aug 2019

    A digital product to get young people reading, built with the people who would use it.

    • Built a mobile-first digital solution to encourage reading among young people
    • Ran co-creation sessions using Design Thinking
    • Worked in Agile Scrum from ideation to delivery
    • Node.js
    • Design Thinking
    • Scrum
    • Mobile

    Carried forward

    • Feniria Labs: Co-creation again: the customer shapes the product.

    Research Support Technician · Universidad Carlos III de Madrid

    Dec 2019 – Apr 2020

    Python for game theory: fair equilibria in complex incentive distributions.

    • Developed Python solutions for game theory problems
    • Researched fair equilibria for complex incentive distribution
    • Coordinated with an international research group
    • Python
    • Game Theory
    • Pandas

    Carried forward

    • Orange: Game theory came back as AI governance and cost routing: incentives, budgets, fair allocation.

    AI Researcher · Universidad Carlos III de Madrid

    Sep 2022 – Aug 2023

    Genetic algorithms for association rules on big data, for customer profiling.

    • Designed custom binary-logic algorithms for association rules on massive datasets
    • Implemented the ML without dedicated libraries
    • Prepared, anonymised and analysed large datasets for cross-sell research
    • Final project graded 9.8 and proposed for publication
    • Python
    • Genetic Algorithms
    • Association Rules
    • Big Data

    Carried forward

    • MásMóvil: Wrangling research datasets was the warm-up for a 5 PB data platform.

  6. Layer 01 · 2012 – 2023

    Foundations

    Math, a double degree, contests, sport and music: the raw features everything else is computed from.

    Lesson · Fundamentals compound

    Trait · Problem solving

    Math olympiads · Escuela de Pensamiento Matemático Miguel de Guzmán

    2012 – 2015

    Competition mathematics and the habit of sitting with a hard problem.

    • Represented IES Diego Velázquez in several math competitions and olympiads
    • Organised the school's first team for Madrid's team math olympiads
    • Trained advanced mathematical and logical thinking at the Miguel de Guzmán school
    • Problem solving
    • Logic
    • Proofs

    Carried forward

    • UC3M: Olympiad problem solving turned into game theory research code.

    Mathematical Engineering · Universidad Complutense de Madrid dropout

    2015 – 2016

    Dropped after one epoch. The network generalised better without it: loved the math, found programming instead.

    • Completed the first year of the degree
    • Loved the math, not the way it was taught
    • Discovered programming and switched paths
    • Math
    • First line of code

    Double Degree, Computer Science & Business Administration · Universidad Carlos III de Madrid

    2017 – 2023

    Engineering and business in one degree, with a specialisation in computing and AI.

    • Graduated with 8.33/10 and honors in Machine Learning, algorithms and game theory
    • Madrid Excellence Scholarship three years running (2019 – 2022)
    • Final project on genetic algorithms for association rules in big data, graded 9.8
    • Student representative of the degree, 2019/2020
    • Machine Learning
    • Algorithms
    • Game Theory
    • Finance
    • Strategy

    Carried forward

    • Grupo SM: Business coursework met Design Thinking in a first real product.
    • UC3M: Honors in game theory became a research job on fair equilibria.
    • UC3M: The ML honors and the 9.8 final project are the same thread.
    • Feniria Labs: Business Administration is why the first pilot became a paying contract.

    Code contests · Google Hash Code · Advent of Code

    2021 – 2024

    Competitive programming for fun, and to stay fast.

    • Top 10 in Spain in Google's Hash Code 2021
    • Top 25 in the world in Advent of Code 2024
    • Python
    • Algorithms
    • Speed

    Carried forward

    • UC3M: Contest-speed algorithms made genetic search over big data feasible.

    High-performance volleyball · Court and beach volleyball

    2020 – present

    Competitive sport as a training ground for discipline and teamwork.

    • Madrid champion in beach volleyball (2021/22) and court volleyball (2022/23)
    • Top 12 beach volleyball team in Spain, season 24/25
    • Four seasons in Spain's second division of court volleyball
    • Discipline
    • Teamwork
    • Recovery

    Carried forward

    • Feniria Labs: Team sport trains what early-stage teams run on: trust and fast recovery.

    Percussion · Symphonic band of Torrelodones

    2013 – 2017

    Classical, jazz and rock, mostly keeping time for everyone else.

    • Member of the symphonic band of Torrelodones, 2013 – 2017
    • Played in bands and orchestras from classical to jazz and rock
    • Rhythm
    • Ensemble

Appendix A

Writing

Notes on AI engineering, agents and the economics of automation. Most of it lives on Medium.

All writing on Medium
Portrait of Hugo Romero
x: the input, before the forward pass.

Appendix B · Contact

What are you trying to solve?

Slow or expensive AI, agents that need to reach production, an AI architecture to design, or a first engineer for an early-stage team. Tell me about it and I will tell you where I can help.

Pick the closest, or describe it in your own words:

Email Hugo

Opens your email app with the message drafted. Nothing is sent from this page.