MedQuery: Evidence-Based Clinical Decision Support

Developed a complete RAG pipeline: retrieved articles from PubMed using NCBI E-utilities, embedded abstracts with sentence transformers, indexed them using FAISS, and generated chain-of-thought explanations with open-source language models.

Enabled explainable, verifiable answers by linking every claim to source articles — bridging the gap between LLMs and evidence-based medicine.

Tools: Python, Streamlit, Hugging Face Transformers, SentenceTransformers, FAISS, PubMed API, Torch

*currently working on deployemnt for easy accessibility

Register-Augmented LLM Finetuning

Collaborated with a team to develop a novel “register-augmentation” technique for transformer-based language models (BERT), improving question-answering performance by adding specialized “register” tokens during fine-tuning.

Implemented interpretability methods (Integrated Gradients, Layer-wise Relevance Propagation) to visualize how register tokens enhance the model's focus on task-relevant context, resulting in better F1 and ExactMatch scores compared to standard fine-tuning.

Answer-Aware Question Generation System

Fine-tuned T5 and BART on QA datasets to generate high-quality questions, using BLEU, ROUGE, METEOR, and BERTScore to compare models and optimize performance.

Ann Arbor Water Production Forecasting

Using Voting Regressor for Time-Series Prediction

Built a predictive model to forecast daily water production in Ann Arbor based on 8 years of weather and usage data. Engineered temporal features and environmental indicators to capture seasonal demand fluctuations driven by rainfall and temperature.

Implemented and benchmarked multiple regression models — including Random Forest, Gradient Boosting, and Ridge and combined them using a Voting Regressor for improved accuracy and stability.

Enabled data-driven water resource planning by predicting future demand and reconstructing sensor gaps, supporting the city’s sustainability goals.

Tools: Python, Scikit-learn, Voting Regressor, Pandas, Matplotlib, NumPy

ThermDepth

Forecasting Subsurface Lake Temperatures for Environmental Intelligence

Predicted hourly temperatures at 10.5m depth in Trout Lake (2018–2019) using multivariate time-series data from 2012–2018 across varying depths.

Engineered temporal and spatial features; benchmarked models including Random Forest, XGBoost, and LSTM for best performance.

Supported climate-aware ecological monitoring by reconstructing missing sensor data with high accuracy.

Tools: Python, Scikit-learn, XGBoost, LSTM, Pandas, Matplotlib

PotholeSense

Predicting Pothole Severity in Chicago Using LightGBM for Smart Infrastructure Maintenance

Built a supervised machine learning pipeline using LightGBM to classify pothole severity across Chicago, utilizing 311 service request data, traffic volume, and road condition features. The model achieved 92% precision and 88% recall in identifying high-risk potholes, enabling data-driven prioritization of repair work

Engineered predictive features including response lag, weather-seasonality, and zip-code-based clustering. Applied robust preprocessing—missing value imputation, label encoding, and outlier handling to prepare the data for training.

Benchmarked LightGBM against Decision Trees, Random Forests, and XGBoost using stratified cross-validation and grid search. LightGBM consistently delivered the best performance with low inference time.

Tools Stack: Python, LightGBM, Scikit-learn, XGBoost, Decision Trees, Random Forests, Feature Engineering, Hyperparameter Tuning, Cross-validation, Pandas, Geopandas, Matplotlib, Seabornb