Projects

Overview

A browser-based Monte Carlo retirement simulator that stress-tests FIRE (Financial Independence, Retire Early) plans against 10,000 randomized market scenarios and four real historical crisis sequences (1929, 1966, 2000, 2008). Users input their portfolio, spending, stock/bond allocation, and age, then see success probability, percentile trajectories, and worst-case drawdown — all computed client-side in under half a second.

Methodology

• Monte Carlo engine with Box-Muller transform for N(0,1) normal random generation
• Cholesky decomposition for correlated stock/bond returns (ρ=-0.05)
• Real (inflation-adjusted) returns sourced from Dimson/Marsh/Staunton and Shiller datasets
• Withdraw-first-then-grow convention matching the Trinity Study / Bengen methodology
• Historical crisis backtests against actual 30-year sequences from each market downturn
• Percentile trajectory computation (5/25/50/75/95) for fan chart visualization

Key Results

Overview

A production-ready FastAPI service that wraps a HuggingFace DistilBERT sentiment classifier behind a clean HTTP interface. Designed as a reference for deploying ML models with the concerns that are easy to get wrong the first time: singleton model lifecycle, Pydantic request validation, batched inference, structured logging, and a multi-stage Docker build.

Methodology

• FastAPI with Pydantic v2 request/response schemas and auto-generated OpenAPI docs
• HuggingFace Transformers pinned to a specific model revision hash for reproducibility
• Singleton model loaded once at application startup via FastAPI lifespan (2–5s load amortized across all requests)
• Batched inference endpoint that amortizes tokenization and GPU transfer overhead (~5x throughput vs sequential)
• Device-aware placement with automatic CPU/CUDA/MPS fallback
• pytest suite covering happy paths, validation failures, and model fallback behavior
• Multi-stage Dockerfile that pre-downloads the model so runtime starts deterministic and offline-capable

Key Results

Overview

A desktop analytics platform designed for proprietary trading firms. Ingests raw trade CSV exports, aggregates fills using interval overlap analysis, and computes per-day statistics including expectancy, profit factor, drawdown, and Sharpe ratio. Models firm-specific payout eligibility rules in real time so traders can track their path to funded accounts.

Methodology

• Interval overlap algorithm for aggregating partial fills into complete trades
• Rolling window statistics for drawdown detection and streak analysis
• Rule engine modeling payout structures from multiple prop firms (FTMO, MFF, TFT)
• SQLite for persistent local storage with migration support
• React component architecture with Recharts for interactive P&L curves

Key Results

Overview

End-to-end classification project predicting whether SpaceX Falcon 9 first-stage boosters will successfully land. The pipeline spans data collection via SpaceX REST API, exploratory data analysis, feature engineering, SQL-based analysis, and an interactive Plotly Dash dashboard for stakeholder exploration.

Methodology

• Data collection from SpaceX API and web scraping Wikipedia launch records
• Feature engineering: orbit type, launch site, payload mass, booster version, flight number
• SQL analysis for aggregation and filtering on launch site success rates
• Model comparison: Logistic Regression, SVM, Decision Tree, KNN with GridSearchCV
• Interactive Plotly Dash dashboard with launch site maps and success rate visualizations

Key Results

Overview

CNN-based medical image classifier for identifying metastatic cancer in 96x96 pixel histopathologic tissue scans. Built and compared multiple deep learning architectures including custom CNNs and transfer learning approaches, with extensive data augmentation to handle class imbalance in the 220K+ image dataset.

Methodology

• Custom CNN architectures with batch normalization and dropout regularization
• Transfer learning with pre-trained models (VGG16, ResNet50)
• Data augmentation: rotation, flipping, zoom, brightness adjustment
• Class-weighted loss function to address label imbalance
• Learning rate scheduling with early stopping for optimal convergence

Key Results

Overview

Binary text classification challenge distinguishing real disaster tweets from non-disaster tweets. Explored multiple NLP approaches from traditional TF-IDF baselines to deep learning architectures including bidirectional GRU and LSTM networks with pre-trained GloVe embeddings.

Methodology

• Text preprocessing: lowercasing, URL/mention removal, lemmatization, stopword filtering
• TF-IDF + Logistic Regression baseline for comparison
• GloVe 100d embeddings with BiGRU and BiLSTM architectures
• Spatial dropout and recurrent dropout for regularization
• Threshold tuning on validation set for optimal F1 score

Key Results

Overview

Photo-to-painting transformation project comparing three generative approaches: DCGAN, CycleGAN, and Neural Style Transfer. Trained models to produce Monet-style artwork from photographs, exploring the strengths and limitations of each architecture for artistic style transfer.

Methodology

• DCGAN baseline: generator/discriminator pair for unconditional Monet generation
• CycleGAN: unpaired image-to-image translation with cycle consistency loss
• Neural Style Transfer: VGG19 feature matching for content/style separation
• Instance normalization and spectral normalization for training stability
• MiFID (Memorization-informed Fréchet Inception Distance) for evaluation

Key Results

Overview

Predicting Rotten Tomatoes critic scores from raw screenplay text using a fusion architecture that combines BERT transformer embeddings with handcrafted numeric features. The model learns semantic patterns in movie review language and maps them to continuous score predictions.

Methodology

• BERT-base tokenization and fine-tuning for domain-specific text understanding
• Fusion architecture: BERT [CLS] embeddings concatenated with numeric features
• Multi-layer regression head with dropout and batch normalization
• Cosine annealing learning rate schedule with warmup
• Evaluation: MAE, RMSE, and R² on held-out test set

Key Results

Overview

Full-stack project management application built for an interior design company. Features task tracking with drag-and-drop Kanban boards, client management, project timelines, and team coordination. Designed for small teams managing multiple concurrent design projects.

Methodology

• React frontend with component-based architecture and React Router
• Node.js/Express REST API with JWT authentication
• PostgreSQL database with normalized schema for projects, tasks, and users
• Role-based access control for admin, designer, and client views
• Responsive design optimized for both desktop and tablet use

Key Results