Experience

2026-01 — 2026-07

Machine learning engineer — agent development

Development of the agent behind an inspection assistant for Red Eléctrica de España. Technicians had millions of photos of high-voltage towers and their defects, and had to review them tower by tower. With the agent they can query by area and by defect type, which makes prioritising and dispatching repair crews far more efficient.

Responsibilities

Design, development and deployment of the whole agent, including its infrastructure.
Integration with the image analysis and vision model training backend built by the vision team.

Key Accomplishments

I redesigned the architecture: it went from detecting the user's intent and running a single tool, to a multi-step ReAct loop over around 10 tools (image analysis, S3 navigation, vision model training, sending reports by email). The agent stopped answering isolated commands and started resolving chained requests end to end, such as analysing towers across several areas of Spain and then sending a report.
I introduced LangSmith for traceability: finding the cause of a failure went from over 10 minutes reading logs by hand to seconds. Most importantly, it made it possible to reproduce exactly the conversation of a colleague reporting a problem; before that took days or weeks and ended in the classic "works on my machine".
I implemented multi-provider LLM support with automatic fallback when a provider goes down, a step limit, guardrails and cache management to cut token spend.
Per-user memory, so the agent adapts to each person's preferences.
Microsoft Teams and web interfaces.

Python

LangGraph

LangSmith

MCP

OpenAI

Mistral

OpenRouter

AWS S3

Railway

Docker

2025-03 — 2026-01

Machine Learning Engineer - Freelance

Alvea Soluciones Tecnológicas Remote

Alvea hired me as a freelancer to build agent products with their partner company, Turing Dream. The main project was a multi-agent tutoring system for secondary school students, with one agent per subject and course material written and validated by a team of teachers. It was used with real students at a private school in Spain and at a secondary school in Colombia.

Responsibilities

Agent task planning, both inside the context window and by generating execution-plan files.
Multi-agent systems and agents with sub-agents, splitting the problem into smaller ones to raise the probability of success without saturating the context window.
Tool development to give LLMs new capabilities: file management, API calls, code execution.
Prompt engineering with positive and negative examples, behaviour conditioning and response shaping.
System prompts that define behaviour, provide problem context and harden the agent against prompt injection.
Human in the loop, so the user can steer the agent mid-task and keep it on course.
Context engineering: adding and removing information, compaction and pruning, splitting context across sub-agents, avoiding context poisoning.
Graph databases and Gremlin queries.

Key Accomplishments

Delivered two MVPs in two weeks.
Took the tutoring agent from prototype to real students in two countries.

Python

OpenAI Agents SDK

Graph databases

Gremlin

2024-12 — 2025-03

Machine Learning Engineer - Freelance

Civir Remote

Development of a chatbot for HoReCa (hotels, restaurants and cafes), through which cooks and waiters can talk to the chatbot to make queries about recipes and menus.

Responsibilities

Backend development with chatbot
RAG development
Implementation of agents that obtain the user's intention and execute actions
Development of React components for the user interface
Implementation of chatbot through spoken communication

Key Accomplishments

Customizable vector searches
Prompt generation with LangChain, built through variables from the user's question and the company's information
Designed a streaming audio pipeline to cut STT cost: the front sent 100 ms chunks, the backend reassembled them, used VAD to gate on actual speech and buffered continuously, sending audio to the STT API only after ~2 s of silence, so only real speech was ever transcribed.

Python

LangChain

PostgreSQL

PGVector

React

Kubernetes

Docker

GitHub Actions

2024-09 — 2024-11

AI tech lead

Bravent Remote

Technical leadership on a generative AI initiative on Azure: architecture, technology selection and risk assessment. The programme was halted by the client.

Python

Pytorch

HuggingFace

Azure

Azure Machine Learning

2022-09 — 2024-09

Machine Learning Engineer

Sener Tres Cantos, Madrid

Machine Learning Engineer, vision algorithms development for autonomous vehicle and leadership in RAG system to obtain documentation information

Responsibilities

Development of the set of vision algorithms for autonomous vehicle.
Leadership in the development of RAG system to obtain documentation information.

Key Accomplishments

When working on an autonomous vehicle everything has to go on an embedded device, which is not as powerful as a normal computer. So I optimized the neural networks with TensorRT, making the inference time and the necessary VRAM memory much lower, ensuring the real-time operation of the autonomous vehicle.
Development of a dataset of realistic synthetic images to train the networks for the autonomous vehicle, creating images in different environmental conditions, making the neural networks more robust.
Implementation of startup scripts for all the libraries and programs of the autonomous vehicle device. Before only one person knew how to install everything, now anyone can start the system.
Development of the complete perception stack (segmentation, classification, detection and depth estimation) for an autonomous vehicle in unstructured environments, within a European defence consortium.
Development of a labelling tool that cut the time to label an image from hours to minutes.
Mentoring of 3 interns and juniors.
Drove the promotion of a colleague from the vision team.

Python

Pytorch

HuggingFace

TensorRT

Nvidia Jetson

2017-10 — 2022-07

AI and electronic engineer

Arquimea Leganés, Madrid

AI, HW and FW development.

Responsibilities

Implementation of detection algorithms for UAV.
Development of a pilot algorithm for geopositioning without GPS.
Leadership of UAV HW and FW development.

Key Accomplishments

With the implementation of detection algorithms for UAV, we avoided having to buy a device that did such detection, saving a cost of €2000 per unit.
When I arrived, the HW and FW of each of the UAVs and ground stations was different, which meant that there were several teams developing the same functionality, in different ways. I unified the HW of the UAVs and ground stations, which reduced the cost of PCB manufacturing. It also allowed to create a common low-level FW, having to change only the high-level part, avoiding duplicate developments.
In electronic warfare environments it is common to disable GPS, which is lethal for a UAV: without positioning it can end up going down in enemy territory and being reverse engineered. I developed a GPS-free geopositioning MVP, training a network on overhead imagery of Spain and France, able to locate the aircraft from a single image with an error of approximately 1 km. That accuracy is enough for the real goal: not to navigate precisely, but to bring the aircraft back to friendly territory, which spans thousands of km². It was delivered as an MVP and never reached production.

Python

Pytorch

YOLO

Altium

STM32

2014-05 — 2017-10

Electronic Engineer

Indra Torrejón de Ardoz, Madrid

HW and FW development.

Responsibilities

FW development of autonomous helicopter control PCB.
Maintenance of autonomous helicopter PCB HW.
HW development of PCB with sensors for a fighter.

Key Accomplishments

Nobody wanted to take charge of the FW of the autonomous helicopter control PCB, which made the project progress very slowly. I took charge of the FW, so the project progressed much faster.
When I arrived, the startup tests of the PCBs were done by hand, since they were designed by the engineers who designed them and only tested them in the design phase. But when the project was in production, every time new PCBs were manufactured, their tests were very slow because they had to be done by hand. In the PCBs I developed, I implemented test wiring and test FW, so that the test of the PCBs was done automatically, making that when the project was in production, the tests of the PCBs could be done much faster and also could be done by more people, because it was not necessary to have personnel with a certain minimum knowledge.

Altium

Eclipse

2013-01 — 2014-05

Intern

Indra Torrejón de Ardoz, Madrid

Help project manager in project management. HW and FW development.

Responsibilities

Help project manager in project management.
Design tests for starting communications PCBs for air traffic controllers.
HW and FW design

Key Accomplishments

Design and documentation of tests for starting communications PCBs for air traffic controllers.
Repair of PCBs for air traffic controllers.

Altium

MPLAB

Explore More

Continue reading

Streaming an agent response: from SSE events to text on screen

How an agent prints the LLM response while it is being generated: splitting SSE events broken up by the network, separating reasoning from the answer, and batching renders so the UI stays responsive.

The agent loop: feeding results back to the model and knowing when to stop

How a harness closes the tool call cycle: it runs the tools, feeds the results back into the history, calls the model ag...

Running Qwen3.8-27B locally on two RTX 3090s: vLLM, MTP and a private API

How I installed Qwen3.8-27B on a PC with two RTX 3090s: vLLM vs llama.cpp benchmarks, MTP, TP vs PP and an on-demand pri...

Last posts -->

Have you seen these projects?

Gymnasia

LangGraph Deep Researcher

Tau

View all projects -->

>_ Available for projects

Do you have an AI project?

Let's talk.

maximofn@gmail.com

Machine Learning and AI specialist. I develop solutions with generative AI, intelligent agents and custom models.

Write me LinkedIn

Do you want to watch any talk?

Tomorrow's Agents: Deciphering the Mysteries of Planning, UX and Memory

AI agents, powered by LLMs, promise to transform applications. But are they simple executors today or future intelligent collaborators? To reach their...

Create your own Apple intelligence

Learn to create an IA system to execute efficiently on a device

Last talks -->

Do you want to improve with these tips?

Best practices building agents with Claude Code

Technical talk: skills, subagents, slash commands and MCPs in Claude Code

o1 prompt engineering

Create better prompts for o1 following an example

Memory profiler

See the memory usage of a script

Last tips -->

Use this locally

Hugging Face spaces allow us to run models with very simple demos, but what if the demo breaks? Or if the user deletes it? That's why I've created docker containers with some interesting spaces, to be able to use them locally, whatever happens. In fact, if you click on any project view button, it may take you to a space that doesn't work.