I build AI systems that turn unstructured data into structured, traceable inputs for financial decision-making.

Mathematical Engineering and Artificial Intelligence graduate (ICAI, Universidad Pontificia Comillas). Master in Finance candidate at IESE Business School, Class of 2027.

backtest-overfitting: the selection-bias corrections of Bailey and Lopez de Prado, implemented and applied end to end. Probabilistic and deflated Sharpe ratio, minimum track record length, minimum backtest length, effective number of trials by correlation clustering, and probability of backtest overfitting by CSCV. On a 147-strategy zoo over SPY from 1995 to 2026, the best in-sample Sharpe of 0.52 gives a deflated Sharpe ratio of 0.58, short of the 0.95 threshold, and no rule in the zoo beats buy and hold. The formulas are checked against Monte Carlo simulation, and the package ships an audit CLI that takes your own matrix of trial returns.

lazy-prices: replication of Cohen, Malloy and Nguyen (2020) on the current S&P 100, built on the 10-K corpus from LUCA. The effect does not survive: FF5+MOM alpha of -0.9 percent a year with t = -0.5, and event-time abnormal returns, within-cohort rank correlations and a 200-run placebo all agree. Of 54 variants, the best reaches an annualised Sharpe of 0.54 and a deflated Sharpe ratio of 0.13, measured with the package above. The write-up is an eleven-page research note whose every number is generated from results/ by the code.

black-scholes-greeks: Black-Scholes prices and Greeks for European calls and puts, checked against an independent Monte Carlo pricer that simulates GBM paths and never uses the closed form. On Hull's textbook example the closed-form call is 4.7594 and one million paths give 4.7580 with a standard error of 0.0050, a gap of 0.28 standard errors. All five Greeks agree within 1.3 standard errors, and the closed form reproduces the values printed in Hull. The tests also check that the Monte Carlo 95 percent intervals contain the exact price about 95 percent of the time, which catches a mis-scaled error bar that price checks alone would miss.

SAM: multi-agent LLM system that automates the full systematic review and meta-analysis pipeline. Nine specialized agents covering protocol design, multi-database search, screening, data extraction, quality assessment, statistical synthesis and manuscript generation, plus a per-paper traceability report. Statistics run in R (metafor), not in the LLM. 879 passing automated tests. Presented at CIPIE 2026.

LUCA: LLM + XBRL pipeline that extracts structured financial data from SEC 10-K filings. Figures come from XBRL and are never generated by the model: 100 percent extraction accuracy on a hand-annotated gold standard of 60 datapoints (6 metrics across 10 companies), against 50 percent for the best LLM baseline. Runs a quantized 7B model on an 8 GB consumer GPU.

NerD: multitask BiLSTM that tags named entities and scores sentiment on news and social posts in one pass, then turns the structured result into a reputation alert.

QUILL: five cooperating LLM agents for automated feedback on student essays (Sogang University, Spring 2026).

DiSpAtCh: Double DQN agent with positive-experience prioritization and transfer learning for warehouse navigation and delivery.

• BSc in Mathematical Engineering and Artificial Intelligence, ICAI (Universidad Pontificia Comillas). Exchange semester at Sogang University, Seoul.
• Master in Finance, IESE Business School, Class of 2027.
• Languages: Spanish (native), English (fluent), German (B2).

• LinkedIn: linkedin.com/in/iqueipopg
• Email: iqueipo.pg24@gmail.com