Built a RAG system for answering clinical questions on diabetes and hypertension using 9 raw medical documents. Cleaned and chunked documents into 200-word segments with 40-word overlap, generated embeddings using MiniLM-L6-v2, indexed with FAISS for fast retrieval, and evaluated GPT-2 family models to reduce hallucinations and improve factual grounding.

Identified regional food insecurity gaps across DC, MD, and VA using Capital Area Food Bank data. Cleaned and structured raw government datasets in Excel, built an interactive Power BI dashboard comparing food distributed vs. unmet demand across regions. Revealed that Maryland had the largest shortfall despite highest distribution — enabling Downtown Evening Soup Kitchen to make data-driven resource planning decisions.

Predicts Tesla stock movement by combining historical stock data and real-time news sentiment. Fetches stock data via yfinance and news via NewsAPI, merges and processes both, trains a machine learning model on AWS SageMaker, and serves predictions through a Streamlit web app. Visual insights are displayed using AWS QuickSight.

Designed an AI-driven healthcare analytics MVP to assist clinicians with real-time risk monitoring and decision support. The system analyzes patient vitals and lab data to generate risk scores, visual dashboards, and automated clinical reports — reducing manual review time. Built as a concept prototype focusing on practical clinical workflow integration.

Analyzed MTA department workforce budgets against actual staffing positions from 2017 to 2023. Cleaned and structured public NY state data in Excel, built pivot tables and KPI tiles tracking total budget, actuals, and variance. Delivered an interactive dashboard with year and status slicers — enabling non-technical stakeholders to identify over- and under-performing departments and make data-driven workforce planning decisions.

Built a transformer-based deep learning model for pixel-level semantic segmentation of leaf health from real-world images. Fine-tuned SegFormer with transfer learning, optimized using Dice loss combined with Cross-Entropy to accurately classify healthy vs dry leaf regions. Evaluated ground truth masks against predicted outputs to measure segmentation accuracy.

Trained two separate TensorFlow CNN models — one to classify floor patches as dusty or clean (85% accuracy), and another to classify stair, plain floor, obstacle, or unknown regions. Integrated both models into a real-time simulation using OpenCV with memory-based movement logic to avoid re-cleaning visited areas. Enabled autonomous multi-floor cleaning by detecting stairs and virtually climbing to continue cleaning across floors.