ARB

An 80-page comprehensive linguistic analysis reframing Italian verb conjugation through the lens of Wittgensteinian "language games" — organizing 15 tenses not by traditional grammatical categories, but by psychological intent families (Facts, Simulation, Subjectivity, Command). Systematically deconstructs each tense into algorithmic components (radical + theme + marker + person ending), maps phonetic evolution patterns (the rolling "R" of simulation vs. the hissing "SS" of counterfactuals), and documents over 200 irregular verb clusters grouped by mechanical pattern. Features detailed conjugation tables, cross-family borrowing analysis (formal imperatives from the subjunctive), and treats "essere" as L'Anomalia Assoluta — the verb so ancient it defies all structural rules.

Proof-of-concept MDP framework demonstrating data-driven approaches to multinational production sourcing under tariff uncertainty. Integrates stochastic trade-war modeling (exponential hazard timing + Negative Binomial magnitude) with dynamic programming and structural estimation techniques (Nested Fixed Point algorithm, Rust 1987) — calibrated to historical US–China trade conflict data. Monte Carlo simulation over 10,000 trajectories validates optimal reshoring strategies.

Comprehensive performance profiling of fine-tuning DistilBERT (67M parameters) on IMDB sentiment classification using PyTorch. Systematically benchmarks training pipeline components including data loading, forward/backward passes, and gradient updates. Analyzes computational bottlenecks across CPU and GPU execution, evaluates optimization strategies (batch size scaling, mixed-precision training), and provides actionable insights for accelerating transformer-based model training workflows.

Full-stack web application for researching US government contracts, contractors, agencies, and spending patterns. Features interactive dashboards with spending analytics and trend visualizations, comprehensive CRUD operations, advanced multi-entity search, and flexible data export. Built with Flask REST API and React TypeScript frontend.

Critical minerals research and analysis framework. Focuses on supply chain dynamics, geopolitical dependencies, and strategic resource assessment for minerals essential to technology, defense, and energy transition.

Deep-dive fundamental analysis of the global copper market, modeling the full mine-to-refined supply chain against sectoral demand drivers. Tracks COMEX and LME inventory dynamics, TC/RC smelting margins, and concentrate market tightness indicators. Quantifies the structural supply deficit thesis through energy-transition copper intensity projections.

Systematic equity-beta analysis of ~15 copper miners against COMEX HG1 futures, combining asymmetric correlation measurement, bull/bear regime decomposition, and risk-adjusted scoring. Features a fully out-of-sample walk-forward backtest with quarterly expanding-window rebalancing — no lookahead bias.

End-to-end analysis of European natural gas markets integrating 10+ data sources. Builds storage-level forecasting models, demand decomposition with HDD/CDD weather features, and TTF price models across the full 2018–2024 pipeline. Tracks Russian route-level flow decomposition, LNG sendout by country, and energy security metrics.

Comprehensive VLCC tanker sector analysis using Frontline (FRO) as the primary equity lens. Includes cross-correlation lead-lag analysis between rates and equities, pairs trading with z-score mean-reversion signals, a Random Forest predictive model, and scenario analysis spanning rate rallies to super-cycle projections.

Quantitative analysis of the gold-silver relationship over ~25 years, from exploratory statistics through a fully out-of-sample walk-forward pairs trading backtest. Applies Engle-Granger cointegration testing, regime identification, and macro correlation sweeps. Integrates Silver Institute supply/demand fundamentals.

Investigation into whether SGX 62% Fe iron ore futures predict one-week-ahead returns for Rana Gruber (RANA.OL). Implements lead-lag validation, FX-adjusted signal construction, and a long/short trading strategy backtest. Demonstrates the gap between statistical and economic significance — a common result in quantitative finance.

Multi-framework quantitative analysis of the XOM/CVX relationship using 55+ years of data. Applies cointegration, multi-factor PCA decomposition, technical crossover analysis, and news sentiment scoring. Key finding: high correlation but no cointegration — 2024 divergence driven by fundamental differences.

Image captioning system combining ResNet-18 encoder with LSTM decoder. Includes beam search, Bahdanau attention with spatial heatmaps, BLEU evaluation, temperature/nucleus sampling, and error analysis. Trained on Flickr8k.

RAG pipeline on SQuAD 2.0 comparing 9 approaches: vanilla, oracle, word overlap, TF-IDF, BM25, dense BERT, hybrid retrieval, cross-encoder re-ranking, and few-shot prompting. Includes Recall@k curves, error decomposition, and confidence calibration.

Five bandit algorithms (ETC, UCB, Thompson Sampling, LinUCB, GLM-UCB) applied to theoretical Gaussian bandits and a real 90K-player Cookie Cats A/B test. Covers regret theory, Bayesian posteriors, and contextual personalization.

2-segment Exponential-Weibull mixture for NYC license processing times with right-censoring, covariate-dependent mixing, hazard analysis, Bayesian updating from approval timing, and multinomial logit outcome prediction.

Discovers two latent cyclist populations on NYC East River bridges via 2-segment Poisson mixture regression. Demonstrates LogSumExp for numerical stability. Fair-weather riders vs all-weather commuters respond oppositely to rain.

High-performance fractal visualization tool generating 1080p videos zooming 100,000x into the Mandelbrot set. Features Numba JIT compilation for near-C performance, psychedelic color cycling, and adaptive iteration scaling.

Educational implementations of k-Nearest Neighbors and Gaussian Naive Bayes classifiers for email spam detection, built from scratch to demonstrate core ML concepts. Trained and evaluated on the Enron-1 dataset.