I'm a Machine Learning Engineer and MS student in Artificial Intelligence at Boston University, with a Bachelor's in Machine Learning. As a Data Analyst at EVERSANA I worked hands-on with large-scale pharmaceutical data, which sparked a lasting interest in applying ML to healthcare and medicine. I build ML systems end-to-end, from data and feature engineering through modeling, evaluation, and the deployed API, across classical ML, deep learning, and LLM systems.
What sets my work apart is a bias toward rigor and internals: I implement algorithms from scratch to understand how they work (PPO in pure NumPy), I build the evaluation and monitoring layers most demos skip (drift detection, LLM-as-a-judge, statistical significance testing), and I hold my systems to measurable standards: grounded answers, verified citations, honest refusal, and reproducible metrics. I'd rather ship one model I can trust and explain than ten I can't.
A clinical RAG that won't lie to a doctor, and proves it.
Local-first RAG over clinical notes with verifiable citations, honest refusal, and grounded extraction.
View project →LangGraph agents that run an entire ML workflow autonomously.
Specialized AI agents that explore, model, evaluate, and explain, coordinated by an LLM super-agent.
View project →An LLM-as-a-Judge framework that measures faithfulness and catches hallucinations.
A reusable evaluation suite scoring grounding and classifying hallucination type per response.
View project →Does an LLM hallucinate more in some languages, and does it depend on the task?
An NLP research study measuring how hallucination rate shifts across languages and task types, with a confound-controlled design and LLM-as-a-judge validation.
View project →Five drift scenarios versus five statistical detectors.
An MLOps toolkit mapping detection methods to drift types and quantifying performance impact.
View project →PPO implemented in pure NumPy, put to work tuning a real classifier.
A from-scratch PPO agent that learns the optimal decision threshold for an imbalanced classifier: RL internals, not library glue.
View project →