ALL PROJECTS
Financial modeling, machine learning, and data-driven research
Built & Live
16 projectsAdvanced Option Pricing Suite
Derivatives pricing engine I built for CQF Module 3 (Exam 2, April 2026). Implements Black-Scholes-Merton, Heston stochastic volatility (priced via the Carr-Madan Fourier inversion), and Monte Carlo with antithetic variates. Reproduces the BSM closed-form to four decimals; measures a 2.9× variance reduction on the European call and 5.6× on the binary.
Exotic Options Lab
Where the Black-Scholes closed form runs out. This page prices Asian, barrier, lookback, and American-exercise options under GBM using Monte Carlo path-dependence and the Longstaff-Schwartz regression. Includes live path-explorer animations and a copy-pasteable Python pricer.
Finite Difference Solver
Solving the Black-Scholes PDE directly on a grid. Explicit, implicit, and Crank-Nicolson schemes side by side, with a live stability visualisation showing exactly when the explicit method blows up past the CFL bound, and projected SOR for American early exercise.
Pricing Where GBM Breaks
A catalogue of markets where lognormal equity assumptions fail and the models that replace them: Garman-Kohlhagen for FX with two interest rates, Vasicek and Hull-White for rates, the Merton structural model for credit, and the Schwartz mean-reverting model for commodities.
Commodity Risk Lab
Applied CQF Module 2.3 risk-machinery (Kupiec POF + Christoffersen independence + combined conditional coverage) to live commodity returns. Built on the daily Supabase pipeline I maintain for Muda Coffee — six tickers, fresh data nightly via GitHub Actions. Each asset gets its own thesis-driven VaR method recommendation; the Cornish-Fisher backtest on KC clears all three tests at the 95% level (p = 0.664 / 0.697 / 0.843).
Spotify Popularity Time-Series Forecasting
Apply the same ARIMA / GARCH stack quants use for asset returns to a non-financial domain: track popularity on Spotify. Built on 586,672 tracks across 72 years (1950-2021). Includes a procedural Web Audio synthesiser that lets you hear what 'high energy, low valence' actually sounds like.
Coffee Futures Snowflake Warehouse
A production-style Snowflake data warehouse for coffee futures research. Galaxy schema (fact constellation) with 8 fact tables (~1.2M rows total) and 19 dimension tables stitched together from FMP, FRED, Open-Meteo, and the CFTC COT report. Designed to make a coffee thesis queryable in one join.
RH-Pipe — Delivery Planning Under Real Constraints
Lakehouse pipeline and multi-truck route planner for a Bay Area furniture delivery operation. Cleans and geocodes raw order files, then plans routes around the constraints the crew actually works under: loading time before departure, LIFO truck packing, lunch at hour four, and a hard 17:00 window. Benchmarked on 159 real orders against the 1-truck FIFO schedule the operation ran by hand: 4 delivery days instead of 11, zero overtime instead of 192 minutes, 74% fewer miles.
Financial Transmission Rights Pricing
Most option pricing models assume the underlying is a stock or a commodity. This paper takes the Black-Scholes framework and adapts it for Financial Transmission Rights, where the underlying is congestion risk on a power grid. Includes jump-diffusion models, seasonality adjustments, and interactive LMP charts.
Eli Lilly DCF Valuation
DCF valuation of Eli Lilly (NYSE: LLY) with pharmaceutical pipeline analysis — revenue projections probability-weighted by drug-approval stage, layered with WACC estimation (5.79% via CAPM on 5-year monthly betas), 9-peer relative valuation, and Monte Carlo sensitivity on terminal value and discount rate. Target $627 vs Mar 2024 reference $579, rated BUY.
MNIST: MLP vs CNN
Technical comparison of a Multilayer Perceptron and a Convolutional Neural Network on MNIST digit classification. CNN reached 99.29% test accuracy vs the MLP's 97.92%, with a tighter generalisation gap (0.56% vs 1.70%). Includes 3Blue1Brown-style architecture visualisations and step-by-step convolution animations explaining why spatial awareness matters.
Customer Segmentation Clustering
K-Means and hierarchical clustering on mall customer data. Identifies 5 spending behavior segments with ARI = 0.942 agreement between methods. The unsupervised methodology generalises directly to regime detection and factor-cluster portfolio construction on the quant side.
Brand Perception NLP — Nike vs Adidas vs Under Armour
Scraped and cleaned ~16K consumer reviews from the App Store, Reddit, and Trustpilot. Ran sentiment (NRC, AFINN, Bing lexicons), LDA topic modelling, TF-IDF, bigram and co-occurrence network analysis, and aspect-based sentiment on price, quality, and sustainability — surfaced the specific language driving each brand's perception per aspect.
Urban Heat Island ML Explorer
Interactive satellite ML classification explorer covering Rio de Janeiro, Santiago, and Freetown. Classifies urban heat island zones using the features extracted by UHI-Pipe.
UHI-Pipe — Satellite ML Feature Library
A Python package I published on PyPI that pulls Sentinel-2, Landsat-8, and Copernicus DEM data from Microsoft Planetary Computer and computes 19 spectral indices (NDVI, NDBI, land-surface temperature, others) into ML-ready DataFrames with parquet caching. The feature substrate behind the UHI Explorer (0.96 F1 on Rio).
Neon Survivor
Arena survival game in a single HTML file. All visuals drawn with Canvas, all audio synthesized with Web Audio API. 6 weapons, 6 passives, 6 evolutions, 7 enemy types, boss fights. Zero dependencies, instant load.
In Progress · Coming Soon
3 projectsAdvanced ML Trading System
Implementation of methods from Lopez de Prado's Advances in Financial Machine Learning. CUSUM filters for event detection, triple barrier labeling, meta-labeling for signal confidence, purged K-fold cross-validation, and Kelly criterion for position sizing. Built in Python with XGBoost.
Amazon Comprehensive Financial Analysis
Full financial analysis of Amazon: income statement forecasting, balance sheet projection, DuPont decomposition, WACC estimation, and DCF modeling with Monte Carlo simulation on terminal value.
ResolvAI
AI-powered IT ticket resolution system built for BNP Paribas ServiceNow. Template-based with 10 pre-configured incident types and automated resolution note generation.