Projects

---

---

Featured Projects

Dementia Detection Using PMC Patients Data

Multimodal clinical AI pipeline using LLMs and Regex feature extraction, achieving 0.894 AUC.

Python LLMs XGBoost

Cost-Optimized Fraud Detection Ensemble

Built an ensemble (XGBoost, LightGBM, CatBoost) with isotonic probability calibration and Optuna Bayesian optimization, reducing fraud costs by 22K-102K.

Python Scikit-Learn Optuna XGBoost LightGBM Probability Calibration

Group vs. Wasserstein DRO for Fair Classification

Formulated Wasserstein distributionally robust classifiers (HDRFC) as LPs in CVXPY, reducing true positive rate disparities by 39% without accuracy loss under distribution shift.

Python CVXPY Optimization Algorithmic Fairness

Activation Intervention Across LLM Architectures

Reimplemented FLORAIN and extended nonlinear activation steering across GPT-2, Pythia, and Gemma-2B-IT, boosting TruthfulQA MC2 scores by 14.3 percentage points via residual stream intervention.

Python PyTorch TransformerLens LLM Steering

Local LLM Flashcard Generator & RAG CLI

CLI tool generating flashcards from Markdown notes using local LLMs (Ollama), FAISS vector search, Pydantic structured outputs, and semantic deduplication.

Python Ollama FAISS Pydantic RAG

Optimization Algorithms in Julia

Modular Julia framework implementing 30+ constrained and unconstrained optimization algorithms (ADMM, Quasi-Newton, Trust-Region) using idiomatic multiple dispatch.

Julia ADMM Quasi-Newton Numerical Optimization

---

Additional Projects & Implementations

Pix2Tex Obsidian LaTeX OCR Plugin

Obsidian plugin and Flask REST API backend converting equation screenshots and clipboard images into LaTeX using vision-OCR.

Python JavaScript Flask Obsidian API Computer Vision

Interactive Neural Network Intuition Tutorial

Interactive Streamlit web app providing real-time visual decision boundary rendering and step-by-step progression from single neurons to multi-layer architectures.

Python PyTorch Streamlit Visualization

Distributed Linear Regression & Matrix Multiplication

Distributed linear regression model implemented in Scala and Hadoop Spark using custom RDD matrix multiplication algorithms.

Scala Apache Spark Hadoop Distributed Computing

Predicting Financial Distress via SEC EDGAR Filings

NLP pipeline using FinBERT and SEC EDGAR filings to extract pre-crisis risk signals 12–18 months prior to balance sheet deterioration during the 2008 crisis.

Python FinBERT Pandas SciPy NLP

Concurrent gRPC Key-Value Store

Java-based key-value store with gRPC client-server architecture, thread-safe concurrent access, and custom error handling.

Java gRPC Multi-threading Concurrent Systems

NBA Web Scraper & Database Pipeline

Python-based web scraper built with BeautifulSoup and PyMySQL to extract, process, and ingest Basketball Reference statistics into a MySQL database.

Python BeautifulSoup MySQL PyMySQL Web Scraping