Data scientist and AI engineer. I work on everything from classic machine learning, computer vision and mathematical optimization to multi-agent LLM systems in production. Mathematician, BSc and MSc in Mathematical Modelling, Statistics and Computation (Universidad de Zaragoza).
At Integra Tecnología I design, build and technically defend AI solutions for enterprise clients in banking, industry, healthcare, public sector and logistics, owning the work from scoping to deployment.
- Multi-agent assistants and conversational systems (LangGraph) deployed in production environments
- Document intelligence: OCR + LLM extraction pipelines integrated with client systems
- Natural-language querying over corporate databases
- GenAI adoption training and mentoring for 30+ companies and public bodies
- Predictive models for business problems: forecasting, classification, regression
- Computer vision systems delivered to clients
- Mathematical optimization for operational and logistics problems
- Data pipelines, SQL and data quality. Previously financial-audit data automation at KPMG (Python, PostgreSQL, Docker, R)
- 2nd prize, HACK THE VIBE hackathon: automatic requirement extraction and eligibility checking over Spanish public tenders
- MSc thesis: LLM techniques for predicting the market impact of financial news
- Trainer on LLM agents and Microsoft Copilot adoption for consulting and industry audiences
Building the public side of this work: evaluation harnesses for agents, retrieval benchmarks, and contributions to the agent tooling I use daily (OpenCode, oh-my-openagent, Hermes Agent).
Python · LangGraph · scikit-learn · pandas · SQL / PostgreSQL · pgvector · FastAPI · Pydantic · Docker · Azure · pytest · Ruff
Zaragoza, Spain · LinkedIn
| catalan-spanish-kindle-dictionary | Free Catalan→Spanish Kindle dictionary: 54.5k lemmas, ~1.1M inflected forms. Data pipeline merging FreeDict, Softcatalà and Apertium with automated format verification. GPL-3.0. |
| energy-forecast-bench | Day-ahead forecasting of Spanish electricity demand: walk-forward backtest over 11 years of hourly data, no data leakage. LSTM beats LightGBM (−0.54 pp MAPE) and closes 44% of the gap to the official REE forecast. MIT. |
| llm-judge-calibration-grid | LLM-as-judge calibration study: 6 judge configurations vs human gold labels on financial QA. Finding: 96% agreement can coexist with zero error recall. Accuracy alone doesn't measure judge quality. |