Built a diagnostics layer for the fund’s existing feature-selection and LightGBM pipeline, now used by the team as a standard model check; it tracks feature redundancy over time (quarterly mean correlation, effective rank, pairwise-correlation drift) to flag the feature space collapsing into fewer independent bets. Made clustering and training decisions auditable: threshold sweeps, cluster-membership churn and representative turnover, plus decile monotonicity, importance by cluster, and the information lost at each stage of the selection funnel. Prototyped an LLM-agent layer on the production model, feeding agents orthogonal top-20/50/100 feature subsets ranked by IC and t-stat, and ran the Lookahead Propensity test (Gao, Jiang & Yan, 2025) to screen for look-ahead contamination. A ~6% out-of-sample IC lift did not survive portfolio construction; jointly recommended against deployment with the PM.
Established the institution’s first student organization dedicated to quantitative finance, scaling membership to 30+ students in its inaugural semester. Teaching foundational quantitative finance concepts to lower-semester members while establishing the framework for the club’s upcoming collaborative group research initiatives. Directed a 5-person algorithmic trading team to finish in the Top 1.5% globally in the IMC Prosperity 4 challenge.
Architected a constraint satisfaction scheduling system for a 150+ user campus environment, formalizing combinatorial resource allocation as a CSP and applying backtracking search with constraint propagation to guarantee feasibility under competing institutional constraints. Engineered PostgreSQL indexing strategies that reduced query latency by 40% under high-cardinality constraint evaluations, and surfaced scheduling state through a real-time React visualization layer.
Built a walk-forward backtester allocating across 10 commodity-sector strategies — 52-week estimation window, 4-week rebalance, Ledoit–Wolf covariance shrinkage, and SLSQP min-variance and max-Sharpe objectives under 40% weight caps — tested over 110 out-of-sample weeks.
Engineered a real-time computer vision system to monitor crop health in Martian greenhouses, utilizing AWS SageMaker for autonomous stress detection. Developed a multi-modal data pipeline to process environmental telemetry and spectral imagery, enabling automated nutrient deficiency diagnosis in extraterrestrial environments.
Originally designed as a predictive ML model for drought patterns in CDMX using hydrological data. Pivoted to a crowdsourced reporting system to bridge data gaps. Designed the data ingestion pipeline to validate user reports against historical meteorological norms.
Accuracy
70% – 85%
5-Day Horizon
Inference
< 500ms
Features
14 indicators
Source
OpenWeather + Crowdsourced
Systemic Risk ModelingData PipelinesCrowdsourcing
Bloomly: Global Bloom Detection System
Oct 2025
@ NASA Space Apps Hack
Engineered a LightGBM-based predictive model leveraging multi-spectral satellite imagery (GEE) and NASA POWER meteorological data. Conducted rigorous dimensionality reduction across 44 distinct ecological indicators to classify global bloom patterns with high precision (AUC/F1 validation).
ROC-AUC
0.72–0.85
F1 Score
0.70–0.82
Features
44 indicators
Source
GEE + NASA POWER
PythonLightGBMRemote Sensing (GEE)
Multi-Agent Simulation
Aug 2025 – Sep 2025
Simulated autonomous agent behavior using Python (Mesa) and Unity to analyze strategic decision-making dynamics. Designed reward-based optimization functions within constrained state spaces, applying Monte Carlo sampling to identify emergent Nash equilibrium patterns.