Our client: Most of the assets they work on resist easy valuation — distressed Spanish loans, repossessed property, the kind of paper the market struggles to price. Our edge is a proprietary corpus of Spanish judicial and financial documents that no foundation model has ever seen, and the machine-learning systems they build on top of it. They are a ~25-person firm that analyzes and advises on Spanish real estate and distressed debt (NPL/REO), and two of those systems sit at the center of the business.

This is a junior role with the reach of a much larger one. As an early, high-impact hire, you would work across both systems — modeling and document AI — as one connected role, not two jobs. What you build shapes the analysis we stand behind and reaches real decisions in days, not quarters. You would work shoulder-to-shoulder with a sharp, multidisciplinary team: engineers, analysts, and specialists in finance and law.

WHAT YOU'LL BUILD

Pricing what almost nobody can price

Predictive and quant models across thousands of Spanish loan and property records — asset valuation, portfolio-performance forecasting, feature enrichment, tabular and time-series modeling. The numbers you produce guide real capital, so evaluation and calibration are where the craft lives.

Teaching a model to read a courtroom

Legal and judicial documents — multi-column, table-dense PDFs where layout and position carry meaning, not flat text — turned into clean, structured data. OCR, layout-aware extraction, NER, and LLM fine-tuning and evaluation. You would weigh the trade-offs that decide whether extraction is dependable or merely demo-ready: OCR + LayoutLMv3 versus OCR-free approaches (Donut, TrOCR) versus vision-language models.

You would own work across both systems end to end — training, evaluation, deployment, and monitoring, in production rather than in notebooks. It is an unusual amount of scope for an early-career engineer, and you will grow into more of it quickly.

WHAT THIS ROLE IS, AND IS NOT

• Real, production machine learning the business runs on — not a research sandbox, and not a thin wrapper around someone else's API.
• Collaborative and consequential: you build alongside talented people on work that matters, not a queue of tickets someone else has scoped.
• Honest about the data: it is genuinely messy, and structuring it is part of the craft — the ambiguity is where the interesting problems live. You will have real infrastructure, experiment tracking, and an evaluation harness to build on.

WHAT THEY OFFER

• Genuine flexibility on where you work: fully on-site in Barcelona or any hybrid split — and the number of days is yours, not a policy.
• A competitive salary for the Barcelona market, discussed openly.
• A real learning budget — books, courses, and conferences.
• A flexible-remuneration plan covering meals and transport.
• The rarest perk: a domain almost no ML engineer gets to touch, and the scope to own it early.

HOW TO APPLY

Their process is short and transparent: an initial questionnaire, 30-minute call with the person you would work with, then two or three conversations and a decision within about two weeks.

Requirements:

• Solid Python and a real grounding in machine learning (PyTorch, scikit-learn, pandas, Hugging Face).
• Evidence that you build and finish things, and that you reason carefully about why a model works or fails.

You do not need years of experience — recent graduates and self-taught engineers are welcome. A degree helps; something real you have built counts for more.

Pluses:

• Document-AI experience (LayoutLM, Donut, TrOCR) or LLM fine-tuning.
• Spanish or Catalan.
• Any exposure to finance, legal, or real-estate data.

Sof-skills:

Curiosity, precision, and clear communication