Cristiana Murgoci

Differential Privacy, Algorithmic Fairness & Cryptography ↗

A self-contained set of lecture notes I wrote for CS 1261, my favorite course at Harvard. Covers differential privacy, fairness impossibility results, the statistical validity of ML claims, and the limits of enforcing fairness at scale.

Neuromail (link coming soon)

An AI-powered inbox assistant that replaces the chaos of Gmail with a focused, one-at-a-time workflow. Reads your inbox, classifies each message by urgency and importance using a GPT-4 API call, drafts replies you can edit, and learns your priorities over time via a scikit-learn ranker. Deployed on Railway.

Systems Software ↗

• Memory allocator: custom malloc/free in C++ with leak detection, corruption reporting, and 60+ test cases
• Binary reverse engineering: defused a compiled bomb by reading x86-64 assembly with GDB, no source code
• OS kernel: built WeensyOS on x86-64: virtual memory, process isolation, fork, copy-on-write, interrupt handling
• Buffered I/O layer: I/O abstraction supporting sequential, strided, and scatter-gather access patterns
• Unix shell: built sh61 in C++ with process management, piping, redirection, and background execution
• Transaction database: concurrent financial transaction system with file I/O and integrity verification

Predicting Formula 1 Lap Times ↗

Built a pipeline to predict F1 driver lap times from pre-race indicators across 271,773 driver-race-lap observations. Merged the Kaggle F1 database with external weather data and engineered features including qualifying pace, constructor, circuit, and driver age. Compared linear regression, polynomial ridge regression with cross-validated regularization, and random forest with early stopping.

Demographic Determinants of Institutional Distrust ↗

Using the 2023 FDIC National Survey of Unbanked and Underbanked Households (69,484 households), we investigated which demographic groups distrust financial institutions and why. EDA revealed stark racial disparities: Black, Hispanic, and AIAN households face unbanked rates 5–7x higher than White households. Logistic regression identified Black race and Midwest region as the strongest demographic predictors of distrust.

The Incentive Pitfalls of Sports Betting ↗

Modeled sportsbook pricing strategy as a game between naive and sophisticated bettors. Naive bettors misperceive probabilities and are swayed by win/loss streaks; sophisticated bettors know the true odds. Simulated three boost scenarios: no promotion, one-time initial boost, and dynamic targeting of discouraged bettors. The simulations showed that dynamic targeting maximizes sportsbook profit by exploiting behavioral biases. Implemented in Python with full agent-based simulation.

Improving Poker Winnings with Game Theory ↗

Argues that pot odds alone are insufficient for optimal poker play and derives the Mixed Strategy Nash Equilibrium for heads-up scenarios. Shows that a GTO player achieves strictly higher expected value than any pure strategy by bluffing at the right frequency (1:2 bluff-to-value ratio at all-in), making the opponent indifferent between calling and folding.