Projects

UW Madison
M.S. in Data Science
University of Wisconsin–Madison  ·  2025–2027 (Expected May 2027)

Agentic ICU Clinical Assistant

Built a clinical decision-support assistant that flags deteriorating ICU patients in natural language. A Claude agent orchestrates two retrieval paths: an MCP server that runs SQL queries over structured patient vitals, and a LangChain RAG pipeline over clinical guidelines stored in ChromaDB. Evaluated with an LLM-judged harness measuring precision, recall, and F1; deployable on Anthropic API or AWS Bedrock. May 2026 (ongoing).

Claude API MCP LangChain ChromaDB AWS Bedrock RAG

Strategy-Selection Steering in Chain-of-Thought Reasoning

Steered DeepSeek-R1-Distill-Qwen-7B away from low-confidence reasoning strategies by probing hidden states at paragraph boundaries in middle-late layers and applying mean-difference steering vectors. Boosted accuracy from ~50% to ~58% on AIME (probe AUC ~0.87). CS 639, Spring 2026.

PyTorch Activation Engineering DeepSeek-R1 Chain-of-Thought Reasoning

LLM Fine-Tuning for Constraint-Following at 1B Scale

Fine-tuned Llama 3.2 1B with LoRA (r=8, 0.007% of params) on RECAST-30K for multi-constraint instruction-following. LoRA lifted Hard CSR from ~28.9% to ~80.0% (5-fold CV, <1% fold variance), substantially outperforming full fine-tuning (~33.2%, which suffered catastrophic forgetting). Built a 9-stage preprocessing pipeline with SHA-256 deduplication and MinHash LSH near-deduplication (Jaccard 0.85), expanding 30K → 90K examples via back-translation and lexical augmentation. STAT 453, Spring 2026.

PyTorch LoRA HuggingFace RECAST-30K MinHash LSH LLM Fine-Tuning

Road Infrastructure Reporting System

Built an end-to-end civic reporting pipeline at MadData 2026 Hackathon. Fine-tuned EfficientNetV2-S on a custom 7,200-image dataset achieving 96.8% test accuracy across 6 road defect classes (potholes, cracks, faded markings, etc.). Integrated a confidence-gated OpenAI API enrichment step that generates detailed complaint descriptions and automatically routes reports to the appropriate city department based on defect classification.

EfficientNetV2-S OpenAI API Python Computer Vision Hackathon
Amrita
B.Tech. in Computer Science Engineering
Amrita Vishwa Vidyapeetham University  ·  2017–2021  ·  GPA: 8.31 / 10

Gesture-based Game Control using a Low-Resolution Camera

This is a final year project. Developed a gesture-based control system using a 3D ConvLSTM model with efficient pre-processing to replace keyboard inputs in games, achieving ~80% accuracy with minimal latency. Evaluated the Gulpio package for caching image frames to address GPU starvation.

Deep Learning TensorFlow Keras ConvLSTM

Web App Testing using Selenium

Automated testing of a faculty dashboard web application using Selenium WebDriver, reducing repetitive test cycles from days to hours. Integrated with Jenkins CI/CD for continuous testing.

JavaScript Selenium WebDriver Jenkins CI/CD Jira

College Admission Management System

Built a web app with admin and user roles. Admins manage colleges, cutoffs, and registrations; users apply for admission and receive status notifications. Deployed via Firebase Hosting.

Python Flask Firebase Pyrebase Bulma

Network Configuration & Troubleshooting

Configured and troubleshot complex network topologies on Packet Tracer and real Cisco hardware as part of CCNA lab training. Gained hands-on experience with routing, switching, and network maintenance.

Packet Tracer CCNA Cisco

SEO Tool to Analyze Live Web Pages

Built an SEO analysis tool that scrapes live web pages and represents keyword frequency distributions in Excel. Leverages word density analysis to understand how search engines rank pages.

Beautiful Soup urllib Xlsxwriter