# Rubén Oliva

> I like pulling apart agent workflows, testing what models can actually do, and improving the tools I use to run them. Oh, and I also like data.

https://arroz.foo

## 01 Current: Data Engineer at Prima.

At Prima Assicurazioni, I work in the Spain Pricing & Underwriting domain. I build and maintain ETL pipelines, automate recurring work, and contribute to internal systems in other parts of the business.

I also take part in an internal AI group that helps drive generative AI adoption across the company.

## 02 Experience: Work history.

Data engineering, cybersecurity training, and AI automation. My recent roles cover all three.

### Data Engineer, Pricing & Underwriting, Prima

Jun 2026 to present · Remote

Building and maintaining ETL pipelines for Spain's Pricing & Underwriting domain, automating recurring work, and contributing to internal systems across the business. Also part of an internal AI group helping drive generative AI adoption across Prima. Python, Databricks, MLflow, AWS.
https://www.helloprima.es/

### Technical Trainer, Cyber Bootcamp Málaga

May 2026 to present · Málaga

Technical Trainer for a national cybersecurity programme run by the University of Málaga's NICS Lab and funded by Google.org. Completed 20+ hours of train-the-trainer sessions, co-designed and deployed a CTF on CTFd, supported the labs, and mentored students.
https://www.nics.uma.es/cyberbootcamp/

### Fullstack Engineer, AI & Automation, Plytix

Mar 2026 to May 2026 · Internship

Built an internal LLM evaluation platform in FastAPI and Angular. It evaluated model outputs against production prompts, ran 80+ models concurrently, and helped select PIM generative AI models. Model cost fell by 30% per use case with no drop in quality.
https://www.plytix.com/

## 03 Projects: Selected projects.

AI systems, software products, and small experiments. Newest first.

### Caret

Jan 2026 to Jul 2026 · Thesis, 10/10 with honours

AI-first collaborative document editor with an agent inside the editing surface. Five containerized services behind an Express gateway, Google OAuth, JWT, and a pgvector RAG pipeline. Deployed with Docker Compose, Coolify, Cloudflare DNS, and GitHub Actions on a Hetzner VPS. Document writing was about 60% faster in testing.

Stack: FastAPI, Express, PostgreSQL, Docker, Supabase, Hetzner, React, TypeScript
https://github.com/arrozet/caret

### Fid

Nov 2025 to Jan 2026 · Kotlin

Fitness app built to make planning and tracking workouts simple. I made it because existing apps asked for too much setup.

Stack: Kotlin
https://github.com/arrozet/Fid

### Basmati

Oct 2025 to Jan 2026 · Team of seven

Calendar platform built by a seven-person team. Seven FastAPI services behind a gateway, plug-in calendar providers, AWS, and a React frontend that keeps UI code out of the domain layer. Highest possible academic grade.

Stack: FastAPI, MongoDB, Docker, AWS, React, TypeScript
https://github.com/arrozet/basmati

### Brain

Oct 2025 · Malackathon, Best Overall Solution

Led four engineers and built an agentic platform for mental-health data analysis in five hours. An orchestrator routes questions to four tools: Oracle RAG search, web search, a Python executor, and a diagram generator.

Stack: FastAPI, LangGraph, Oracle Autonomous Database 23ai, React, TypeScript
https://github.com/arrozet/brain

### Ticket Splitter

May 2025 to Jun 2025 · AI-assisted receipt splitting

Receipt and bill splitter that uses Gemini for OCR and item extraction. Groups can assign items, account for taxes and tips, and split the final cost.

Stack: FastAPI, Astro, Tailwind CSS
https://github.com/arrozet/TicketSplitter

### Mercedes-Benz Data Science Challenge

Apr 2025 to Jun 2025 · 3rd prize

Model for predicting when a customer buys their next car. Adding wishlists and damage history roughly doubled explanatory power over a linear baseline. Final mean absolute error: 165 days.

Stack: R
https://arrozet.github.io/RetoMercedes/

### Cuby

Mar 2025 to Jun 2025 · JavaScript

Small platformer about a black-and-white cube. Built around one mechanic with no decorative elements.

Stack: React, JavaScript
https://github.com/arrozet/Cuby

### Dr. Artificial

Mar 2025 · Dedalus Datathon, 4th place

RAG system over medical data, built in ten days. Prompt restructuring reduced token usage by 98% with no loss in accuracy.

Stack: Flask, JavaScript
https://github.com/arrozet/dr-artificial

### AImbot

Nov 2024 to Dec 2024 · Computer vision

Real-time YOLO detection connected to a mouse controller. Benchmarked on FPS and inference time.

Stack: Python
https://github.com/arrozet/AImbot

### MiniPIM

Oct 2024 to Feb 2025 · C#

Product-information system built for a requirements-engineering course. The model covers entities, roles, and catalogue rules.

Stack: .NET, C#, MySQL
https://github.com/arrozet/MiniPIM

### DAJER, Water management

Oct 2024 · I Malackathon · 2nd, Accessibility

Water-management dashboard with reservoir maps, queryable levels, and a 12-month forecast. Built in one weekend.

Stack: Oracle Autonomous Database, Hugo, JavaScript
https://github.com/arrozet/malacaton

### Feed The Leader

Mar 2024 to Oct 2024 · C#

Clicker game about feeding a leader. The first project I shipped that people played.

Stack: Unity, C#
https://github.com/arrozet/FeedTheLeader

## 04 Competitions: Hackathons.

Five competitions, four prize-winning finishes, and a lot built against the clock.

### Cursor Build Night Málaga, Cursor

Jul 2026 · 4th place

Four-person team. 105 minutes. Traffic-jam challenge.

### II Malackathon, University of Málaga

Oct 2025 · Best Overall Solution

Agentic platform for mental-health data analysis, built in five hours.

### Mercedes-Benz Data Science Challenge, Mercedes-Benz

Apr 2025 to Jun 2025 · 3rd prize

Vehicle purchase timing model.

### Dedalus Datathon Andalucía, Dedalus

Mar 2025 · 4th place

Clinical summary assistant built for the Dedalus Datathon Andalucía.

### I Malackathon, University of Málaga

Oct 2024

First edition. Reservoir data and a 12-month forecast.

## 05 Education: Education.

BSc in Software Engineering. Ranked 1st of 100+ students with a 9.20 / 10 grade.

### BSc, Software Engineering, University of Málaga

Sep 2022 to Jul 2026 · 9.20 / 10 · 1st of 100+

Earned the highest academic distinction in 50% of the degree credits. That amounts to 18 subjects.

## 06 Ways of working: How I work.

My default is specification-driven development. I define the problem, constraints, interfaces, and acceptance criteria before implementation starts.

Coding agents help me explore options, implement changes, and review the result. I choose both the agent and the model for the task, then verify the result (for now ¯\_(ツ)_/¯).

The agent can write code, but responsibility stays with me. I own the trade-offs, the review, and what ships.

### Rules for work that has to ship.

Six checks I use to turn a written brief into software I can stand behind.

#### Decisions need evidence

For choices that affect cost, quality, or latency, I test on work that resembles the real workload.

#### Prototype and product are different

Experiments can stay lightweight. Systems that ship need repeatable builds, useful checks, and a clear release path.

#### Choose the stack for the problem

I have familiar tools, but I do not force one stack onto every project. The architecture should fit the problem.

#### Keep a human responsible

I use agents for exploration, implementation, and review. The brief, trade-offs, and final result still have a human owner.

#### Assign ownership

Someone should know why a system exists, how to judge it, and when it is safe to change.

#### Use metrics for a reason

When a question can be scored, I benchmark it. Otherwise I use a review, a failure mode, or feedback from the person using the system.

## 07 Links: Elsewhere.

[object Object]

- GitHub: https://github.com/arrozet
- LinkedIn: https://www.linkedin.com/in/ruben-oliva/
- Blog: https://blog.arroz.foo/