Madrid, Spain
I'm a software engineer who builds AI products end to end.
LLM agents, computer vision, and the backends, integrations and web apps around them. Forward Deployed AI Engineer at Gambooza and published researcher on LLM agents.
Pour a pint and see what the till misses.
Gambooza puts cameras over the beer taps of bars and restaurants and checks every pour against the till.
I own projects end to end: the meetings with clients and POS vendors, the backend and the integrations, the client dashboard, the company website and the models that run every night in production.
Forward Deployed AI Engineer at Gambooza
May 2026 – now
Bar owners ask on WhatsApp, by text or voice note, and the agent answers from their own data: a table of last night's sales, a chart of the week, the weekly report as a PDF or the video clip of the pint that was never charged.
Simple questions go through fixed workflows and open ones through the LLM. Every client gets its own container and can only see its own venues.
Built with OpenClaw, Pydantic AI and Docker.
“How did the venues do yesterday? Make me a table.”
Sales by venue
Yesterday vs. the same day last year
| Venue | Yesterday | 2025 | Δ |
|---|---|---|---|
| Downtown | €3,180 | €2,990 | +6% |
| North | €2,640 | €2,720 | −3% |
| Riverside | €1,910 | €1,705 | +12% |
| Total | €7,730 | €7,415 | +4% |
€7,730 across the three, 4% up on the same day last year.
How many beers went uncharged yesterday?
The taps poured 412 and the till charged 398. Most of the 14 missing were at bar 2, between 23:00 and 00:30.
Tap 3 opens, a pint is poured, nothing is rung up.
Send me the rest too
Send me last week's report
Pints poured per day
Week 41, all venues
Saturday was the busiest day: 1,240 pints, 14% more than the week before.
0.998
Per-venue YOLO models, measured on days held out for testing, with regions of interest that follow the tap when someone bumps the camera. The method now runs in more than 10 venues.
18
A multi-tenant pipeline on Cloud Functions and Firestore that ingests 72,000 to 180,000 sale lines a week from 6 POS systems and reconciles them with what the cameras saw.
1 min
The client dashboard I built took us off Looker Studio and spreadsheets. It reads straight from Firestore and updates itself, so we check the reports and send them in about a minute.
−35%
I redesigned the video pipeline: full-frame HD storage, rotation and cropping on the GPU, and audits that keep an eye on recall every night.
BDP, Ágora, Glop, Square, Revo and Madisa, each with its own adapter. Revo with OAuth and PKCE, through the vendor's certification process.
Per-venue and group views, benchmarks against the median venue and a sending workflow for every weekly report. It's what took us off Looker Studio.
Designed from scratch and built by me: a completely new site, off WordPress and onto Firebase Hosting, with Playwright tests in CI and a design system the whole team now builds on.
Before Gambooza I spent a year and a half in the MISO group at Universidad Autónoma de Madrid, building agents that test other chatbots.
Research Software Engineer at MISO group, Universidad Autónoma de Madrid
Dec 2024 – Apr 2026
Hi! What can you do?
I can show you the menu, take your order or tell you our opening hours.
Show me the menu, please.
Margherita, pepperoni or four cheese. Which one, and what size?
A large pepperoni.
Anything to drink?
A cola.
That's €14.50. Shall I confirm the order?
5 functionalities found
5 functionalities · transactional chatbot
A full-stack platform I designed and built. It runs TRACER, the lab's Sensei user simulator and its correctness checks as one testing pipeline, and I presented it as a tool demo at ICSE 2026.
Along the way I also wrote Chatbot Connector, a Python library that lets the same tests talk to chatbots built on Rasa, LangChain or the OpenAI API, released to PyPI with GitHub Actions.
Personal, open source
Close to 3,000 downloads
An open-source Obsidian plugin that follows the links between your notes and exports them as one Markdown file, ready to paste into an LLM or print as a PDF.
Pick a note and how deep to follow its links. Each number is how many links away that note is.
Pick a note to export its context.
Client work
A platform for a vehicle engineering consultancy that moved camper homologation out of email threads and into one guided workflow, with roles for clients and engineers.
Client work
A desktop app for the same consultancy that writes the official engineering documents for vehicle modifications in Spain, replacing repetitive manual paperwork.
I like building whole products, from the database to the interface. Lately most of them have AI inside.
I studied Computer Science Engineering and then Data Science at Universidad Autónoma de Madrid, with a semester at the University of Helsinki and a spring school at Northwestern Polytechnical University in China. I was awarded Excellence Scholarships by the Community of Madrid and by the university, and highest honors in Recommendation Systems, Data Structures and Mobile App Development.