I build production-grade ML systems at the intersection of quantitative finance and deep learning, from Physics-Informed Neural Networks for options pricing to Deep RL agents for market making and end-to-end MLOps pipelines.
I'm a Machine Learning Engineer specializing in production ML systems and applied deep learning. I graduated Magna Cum Laude from Auburn University with a BS in Computer Science and an undergraduate certificate in AI Engineering.
My work spans the full ML lifecycle, from model architecture and training to deployment, drift detection, and monitoring. I build systems designed to hold up in production, not just in a notebook.
Physics-Informed Neural Network (PINN) with a [128, 256, 256, 128] Tanh architecture that enforces the Black-Scholes PDE directly in the training loss via automatic differentiation. Trained on real AAPL options data from yfinance with a curriculum that ramps the physics weight from 0 to 0.01 over training. Computes all five Greeks (Delta, Gamma, Vega, Theta, Rho) via PyTorch autograd using stock-price-augmented training data. Prices American options via Least Squares Monte Carlo with Laguerre polynomial basis functions across 10,000 simulation paths, and uses a PPO agent for dynamic delta hedging. Includes arbitrage detection across put-call parity and butterfly spread convexity checks.
Custom Gymnasium environment with 12-dimensional observations and a 2-D continuous action space for bid/ask offset placement. PPO agent trained for 500,000 timesteps using an Avellaneda-Stoikov inspired reward: per-step PnL change plus a quadratic inventory penalty and a terminal liquidation penalty. Models order flow with a Poisson arrival process (λ = 10 orders/second), VPIN-based toxicity tracking, and a square-root market impact function with adverse selection. Hard inventory limit of 1,000 shares enforced at execution. Transaction costs: 1 bps fee plus 0.5 bps slippage. Policy and value networks: MLP (64→64) with Tanh. Expected benchmarks: Sharpe 0.5–2.0, fill rate 40–80%, spread capture 30–60%, max drawdown <5% of initial capital.
End-to-end MLOps platform built on the IBM Telco Customer Churn dataset (7,043 customers, 20 features, ~26% churn rate). Trains an ensemble of XGBoost, LightGBM, and Random Forest models with Optuna hyperparameter optimization (30 trials, 5-fold cross-validation) and SMOTE for class balancing. Achieves ~0.85 ROC-AUC and ~80% accuracy on a held-out 20% test set, with inference latency under 100ms per prediction. Features a SQLite model registry with versioning, Kolmogorov-Smirnov drift detection (p < 0.05 threshold), A/B testing at 95% confidence, and SHAP-based feature explainability.
Production-grade portfolio optimization across a 10-asset universe (SPY, QQQ, IWM, EFA, EEM, TLT, IEF, GLD, SLV, VNQ) combining Modern Portfolio Theory, Black-Litterman Bayesian posterior estimation, and ML-driven return forecasting. XGBoost model trained on 100+ engineered features (rolling returns, volatility, moving averages, RSI, volume) with an 80/20 train/test split. LSTM: 2-layer, 128-unit recurrent network on 60-day input sequences, trained with MSE loss and Adam optimizer. ML-Enhanced mode blends 70% historical signal with 30% ML predictions. Fama-French 5-factor model (Market, SMB, HML, RMW, CMA) for risk decomposition. FinBERT sentiment analysis on a 30-day Finnhub news window. Risk constraints via VaR/CVaR at 95% confidence; Monte Carlo stress testing with up to 50,000 paths over 5-year horizons. Rebalancing triggered at 5% weight drift with 10 bps transaction costs. Models pre-trained on H100 GPU.