Jorge García

// education

Universidad Politécnica de Cartagena (UPCT) Sept 2024 – June 2028

BSc in Data Science and Engineering · Cartagena, Spain

Distinction (Matrícula de Honor) in Non-Linear Optimization, Data Structures, and Programming Fundamentals.
Relevant coursework: Probability, Statistical Inference, Stochastic Processes & Time Series, Linear Algebra, Optimization I–II (Non-Linear), Algorithm Analysis & Design, Machine Learning I–II, Deep Learning.
First Prize, UPCT Reto Fest: co-developed CiceronIA, an AI assistant for judging debate competitions via multimodal analysis (FastAPI, LangChain).

// experience

Lead Backend & Data Engineer Mar 2025 – Present

B2B Management & Analytics ERP (startup in development) · Cartagena, Spain

Designed the backend of a B2B ERP around a universal, industry-agnostic fixed-table data model (PostgreSQL/SQLite) and built its RESTful FastAPI layer supporting real-time BI integrations.

// technical projects

Multi-Asset Anomaly Portfolio Engine

Python, cvxpy, NumPy, Pandas · github

Architected an open-source quantitative backtesting engine exploiting cross-sectional momentum anomalies across the S&P 500 universe, with a Markowitz mean-variance optimizer built on cvxpy.
Resolved the dimensionality curse (N > T) with Ledoit-Wolf covariance shrinkage; diagnosed a units mismatch (z-score alphas vs. daily covariance) that silently collapsed the optimizer into a degenerate 4-stock corner portfolio, and rescaled the objective to annual units.
Enforced T+1 execution, start-date universe construction, and forced liquidation of delisted assets; audited the data layer for survivorship bias and reported the honest negative result.

Baseline — ATP match outcome prediction

Python, scikit-learn · github · live demo

Built an end-to-end ML pipeline over 15 years of ATP data; designed a custom dual-track Elo rating system (overall + per-surface) with importance-weighted K-factor and monthly-prorated inactivity decay.
Engineered a “panic” index quantifying performance under pressure from break-point conversion and match-duration data; validated every feature with walk-forward yearly folds to prevent temporal leakage.
L2-regularized logistic model beat the Elo-only baseline on Brier score (0.2245 → 0.2156) and accuracy; deployed as a live Streamlit app with a public write-up.

Silo — satellite alternative data research

Python, Google Earth Engine, statsmodels · github

Built an alternative-data pipeline estimating flooded rice-paddy area in the Guadalquivir marshes from Sentinel-2 imagery, using combined NDWI/MNDWI water masking and SCL cloud filtering.
Tested the satellite signal against a rice producer’s revenue and EBITDA with three OLS specifications (levels, first differences, trend-controlled); documented small-sample pitfalls rather than reporting a spurious result.

// technical skills

languages Python (advanced), SQL, LaTeX

libraries & tools NumPy, Pandas, scikit-learn, cvxpy, statsmodels, FastAPI, PostgreSQL, Git, Streamlit, Google Earth Engine

concepts Probability, Convex Optimization, Covariance Shrinkage, Walk-Forward Validation, Time Series Analysis, Backtesting & Bias Control (survivorship, lookahead, temporal leakage)

// activities & interests

languages Spanish (native), English (C1 — Duolingo English Test 125)

speedcubing official WCA competitor — 3x3 single 13.76s, average 17.11s

other tennis