projects

Built a two-stage NL-to-SQL pipeline with a fine-tuned Qwen 2.5 generator and Phi-4 validator, benchmarking 9 models across prompting strategies on EHRSQL 2024 to improve clinical query reliability.

Part of the Capstone Thesis under Prof. Lipika Dey, Spring 2025.

Thesis titled: “Dual-Model Framework for Natural Language to SQL Translation in MIMIC-IV Healthcare Databases”

Built and evaluated ensemble machine learning models (Random Forest, XGBoost, LightGBM, AdaBoost) to predict sepsis patient ICU outcomes (in-hospital mortality and ICU length of stay), with careful handling of class imbalance and feature selection, and developed an interactive visualization tool to help clinicians interpret model predictions.

Part of the Capstone Project under Prof. Lipika Dey, in collaboration with Maanas Kejriwal, Fall 2024.

Project titled: “Data-Driven Exploration: Unlocking Predictive Potential and Insights into ICU Outcomes in Sepsis”

Built a multilingual audio-first interface that takes spoken queries, translates between local languages and English, and returns accessible medical information with a focus on low-friction use for non-English users.

Designed and built a mobile robot using Arduino Uno + Raspberry Pi with sensors and a camera to navigate fields, capture soil moisture readings, and estimate crop health via “green percentage” analysis along its path.

Built an N-body simulation in Python to model planetary and comet trajectories, analyze orbital properties and interactions, and used the framework to test (and argue against) the Nemesis hypothesis via observed orbital instability in the simulated system.

Built a working prototype glove that emits sound pulses, measures the echo return time to estimate obstacle distance, and guides the user with beeps that increase in frequency as objects get closer, usable even in low light/dark environments.