José Ángel Carretero Montes

Conversational financial assistant built for the Cátedra Unicaja challenge (University of Granada): natural-language questions about your transactions with LLM-generated charts, and operations (Bizum transfers, investments) that never run without your approval.

• LangGraph graph that separates deciding from executing: the LLM only proposes, a deterministic step checks funds and limits, the graph pauses until you approve, and execution happens in a single idempotent transaction.
• No free-form SQL: the agent picks from 18 tools, each a parameterized SQLAlchemy query, which rules out SQL injection by design.
• 86.4% accuracy in an evaluation on the real pipeline, checked against values computed with direct SQL; ~4.2 s median full response time.
• Nightly proactive dashboard, user memory (RAG), fully local voice mode, and 56 pytest tests.

C++17 implementation and experimental comparison of 14 metaheuristics (Greedy, local search, ILS, simulated annealing, genetic and memetic algorithms) to build investment portfolios that maximize return against risk.

• Evaluated on the IBEX 35, S&P 100, and S&P 500 (up to 457 assets): trained on 2015-2024 data with out-of-sample validation on 2025, 50 seeds per algorithm.
• The memetic AM-Best algorithm achieves the best training and test results on all three indices.
• O(N²) evaluation parallelized with OpenMP and optimized builds (up to ~40× faster), with multi-platform CI on GitHub Actions.
• Parameter sensitivity studies (BLX-α's α, simulated annealing temperature and neighborhood) documented in three technical reports.

Research applying immunology's Danger Theory to early detection of currency crises: a genetic algorithm evolves agents that learn their own rules from market stress signals.

• The agent predicts "Black Wednesday" (the 1992 sterling crisis) one day in advance, relying almost entirely on Bank of England intervention and unemployment, without looking at the price.
• Validated with a synthetic "value trap" test and by transfer to the 1994 Mexican peso crisis, where it recalibrates its threshold, avoids false positives, and predicts the collapse.
• Academic paper and interactive book published on GitHub Pages, co-authored and supervised by Dr. José Luis Sáez Lozano (University of Granada).