Automation and data analyst at Cabify in Madrid, with an industrial engineering background. I turn manual operations work into systems that run on their own: n8n agents with Claude, Python and SQL pipelines, forecasting models and internal tools, measured in hours removed, errors avoided and cost per run.
I studied industrial organization engineering and ended up where operations meet software. I joined Cabify in Madrid in November 2025 as a Customer Operations trainee, moved to Operational Excellence in March and became the team's automation analyst in May.
The job is easy to describe: find a process people repeat by hand, build the system that does it, and keep it running. Most of it is n8n with Claude for the reasoning, Python and SQL where it has to be real software, and small models I train myself when an LLM would be overkill. Everything on this page runs, or ran, in production with real users, and each case study says what broke as well as what worked.
I also teach. I was selected as one of the company academy's n8n trainers, and my first session is on 30 September 2026.
AQM went to the global Customer Operations all-hands in April 2026. In September, AQM, DQM and booking a ride from WhatsApp were part of the Spain block of the company All Hands. AQM is built and shipped; DQM is the one I am delivering now.
Fifteen systems, most of them in production at Cabify Spain. Every card opens a full case study with the architecture, screenshots and what I would change. Screenshots show the real tools rendered with synthetic data; client, supplier and people names are left out on purpose.
Continuous AI audit of the vinyl wraps on the Spanish fleet: vinyl type and condition, damage and plate, from photos taken by drivers and wrapping suppliers. 97% accuracy against human review, 30 to 60 minutes from photo to verdict.
Mac miniOne 0 to 100 quality score per driver, compared only with drivers in the same city, fleet type and tenure, with a 28-day memory and an escalation ladder attached. I lead the build; the simulator already runs on synthetic data.
Book a ride without downloading the app: scan a QR, have a conversation, get a price, get a car. The agent took the company's first ride booked, tracked and invoiced with no human in the loop.
Two documents arrive by email from two different senders. Two Claude agents read them, a sheet cross-checks them, and the ride is created with its driver pre-assigned in a single API call.
Public support agent in the rider Help Center. It resolves sign-in and registration problems on its own and opens a structured ticket only when a person is really needed.
A rules engine and approval portal that decides which commercial products each vehicle carries. Every approved change becomes a rule with an author, a reason and an expiry, so the daily audit enforces it instead of undoing it.
Mac miniDrop a CSV in a Drive folder and the Mac mini sends it to the right internal RPA, waits for the real result and reports it in Slack. It recognises 33 RPA products by their columns and splits heavy files so they run in parallel.
Mac miniA web board that rebuilds itself every morning with the KPIs of the weekly operations review, their history, a forecast and a status that says what is worth discussing. Monday alerts go to each process owner.
The team's shared LLM key ran out of budget. I traced 96% of the spend to one unfiltered agent tool, made each run 63 times cheaper and got prompt caching into the company's n8n model node.
The engine under the fleet audit. CNN and OCR models I trained on 16,000+ labelled photos handle the repetitive checks, and the LLM is kept for judgement calls. Inference cost dropped by about 80%.
Forecasts of support volume, ride requests by city and quarterly KPI targets, plus an Erlang C and CP-SAT shift planner. The volume model cut the error of the official forecast by 58%.
A scheduled scanner that tells the operation what is about to move a city before it moves: matches, concerts, fairs, strikes, closures. It feeds the forecast the one thing history cannot predict, the exception.
Mac miniA read-only copilot that watches my work tools, drafts replies in my own voice and keeps a local memory of decisions and commitments. It proposes; a person always sends.
Claude with my projects, a persistent knowledge folder, custom skills and MCP connectors to Slack, Gmail, Calendar, Drive, Tableau and ClickUp. Triggered by me, not running in the background.
Product Hackathon build with a team of three: a live fleet portal with a unified driver quality score, a decision-impact simulator and AI voice coaching for drivers.
Tremendous work, team! I have no doubt that what is being built here will become the master process that lets us deliver a superior service and stand out on quality control in the sector. Congrats, let's go!
I'm blown away that you built all of this. Congratulations!!
Great progress on this battle: rides can now be cancelled, usage is rate limited, all the data is being stored. Thanks Mat, great work.
Team, we have completed the first ride requested through WhatsApp, handled 100% automatically through n8n. Congratulations, Mat.
Congratulations to Mat on the great work. It is an important step that will let us reduce errors and keep tightening control over the quality of our vehicles.
First of all, a huge thank you to everyone involved, but above all to Mat.
Top. All the work it took to get here. Thank you!
Wow, tremendous! This pilot looks amazing. Keep us posted!
Congratulations on the great work, Mat!!
A thousand thanks for the detail, and congratulations on such a creative and powerful solution!
From the local team, we never get tired of showcasing Mat's outstanding work on the vinyl audit analysis.
This work is top quality. It is going to help so many different teams. Congratulations!
From internal Slack channels in 2026, translated from Spanish. Names are withheld and titles are shown as the team or level of the person.
n8n (advanced)Claude API · tool use · caching
Claude CodeMCPMulti-agent pipelinesPrompt designRAGAgent guardrailsLangChainLiteLLM gatewayOllamaWhatsApp Cloud APISlack APIZendeskPlaywright RPAYOLOv8PyTorchKeras
TensorFlowscikit-learnProphet · NeuralProphetLightGBMSARIMAXEnsemblesOR-Tools CP-SATErlang CGoogle ColabPythonSQLPostgreSQLAmazon Redshift
JavaScriptApps Script web appsTableau + REST API
Looker StudioRedisGrafanaGitGitLab CImacOS services · launchdLovableGoogle Workspace
An 85-page reference I wrote for myself and shared with colleagues: the layers of an agent system, the tool-calling loop in forty lines of Python, when a workflow beats an agent, local versus frontier models, evaluation, security and cost. A personal document, not a company publication.
Read the manual
The workflows behind the case studies, drawn from their real n8n exports: node by node, relabelled in English, with client names, hosts and credentials removed. For when you want to see how the systems are wired rather than read about them.
Open the dossierI'm relocating in October 2026 and looking for a team where automation and AI create real operational leverage: AI operations, automation, data analysis or operational excellence. If that sounds like yours, let's talk.
