Projects
A list of projects that I have been working on or built but not limited to.
Full-stack quantitative research platform combining a multi-factor equity scoring model (momentum, technical, fundamental, PEAD/event factors, cross-sectionally normalized) with a backtesting engine, regime filter, and live paper-trading execution across 500+ tickers via Alpaca API.
Adaptive regime-switching trading system that classifies market conditions into 4 states (bull, choppy, high-vol trend, crisis) using a Gaussian Hidden Markov Model trained on 5,032 trading days. Dynamically allocates across a momentum/mean-reversion/trend/defensive strategy pool via HMM posterior probabilities — achieving OOS Sharpe 0.769 vs 0.641 for pure momentum with max drawdown reduced from -33.7% to -25.6%. Live paper trading on Alpaca with automated weekly rebalancing.
Global DeepAR model jointly trained across a 2,001-series stratified subset of the M5/Walmart retail dataset, producing P10/P50/P90 quantile forecasts (WQL/pinball loss) and evaluated against seasonal-naive, ETS, and Prophet baselines on a rolling-origin backtest. A leakage-tested cold-start holdout — verified by a test that walks the actual training dataloader and asserts no batch can see pre-cutoff data — shows cold-start series performing on par with warm-start (WQL 0.460 vs 0.466). Forecasts feed a business simulation where a P90 order-up-to inventory policy cuts the stockout rate from 23.9% to 1.9% at ~6x the holding cost of a P50 policy.
Configurable two-sided marketplace simulator with a known ground-truth causal effect, built to quantify exactly how naive A/B testing breaks under spatial interference — bias grows from statistically indistinguishable-from-zero at no spillover to -36.6% at high spillover. Validates switchback and buffered cluster-randomized designs that stay within ~5% bias regardless of spillover strength, applies diff-in-diff (cutting a +1,471% naive-comparison bias to -8.5%) and synthetic control to a non-random rollout, and closes the loop by benchmarking offline IPW policy evaluation against true online experiment results across multiple logging-data seeds.
Complete data pipeline processing 400K+ monthly BayWheels bike-share ride records using Python, Prefect, Docker, and GCP. Terraform provisions cloud infrastructure, dbt handles data modeling, and BigQuery powers analysis and storage; Kafka and Spark handle batch and real-time data to track ridership trends, station usage, and trip patterns. Cut manual work by 90% and made reports 60% faster.
Interactive Streamlit app that compares European option pricing using Monte Carlo simulation and the Black-Scholes formula. Includes GBM-based path simulation, real-time pricing, sensitivity charts, and mathematical documentation.
A dynamic financial web app that filters and ranks S&P 500 stocks based on key factors like valuation, profitability, growth, and risk. Includes real-time scoring, data visualizations, and a savings-based investment allocator.
High-performance RAG application built with LangChain, OpenAI, Cohere, Pinecone, Flask, and FastAPI, enabling contextual Q&A over academic PDFs for FSU's Applied Machine Learning course. Reduced research lookup time by 70% through real-time document retrieval and LLM-powered summarization.