NSF-ACCESS CiDeR GenAI Tool

During my summer internship through the NSF-ACCESS STEP program at UIUC/NCSA, I developed CiDeR GenAI — an automated auditing pipeline that keeps ACCESS cyberinfrastructure resource records synchronized with Resource Provider documentation. The core problem: CiDeR database entries were manually maintained and frequently drifted from provider guides that change continuously. I integrated the CiDeR Operations API to fetch live records, built a configurable breadth-first web crawler and documentation scraper to collect provider content at scale, and deployed an LLM auditor on JetStream2 that compares each database field against scraped documentation. To eliminate hallucination risk in structured reporting, I used Pydantic schemas to constrain all LLM outputs to strict JSON — including proposed corrections, supporting evidence, and field-level modification recommendations. I also investigated the ACCESS-MCP server architecture, identified critical data synchronization gaps, and delivered findings to the operations team. The project was delivered as a fully packaged, reproducible Python MVP with documentation for seamless handoff to operations staff.

NSF ACCESSJetStream2CiDeR Operations APILLMStructured OutputPydanticPrompt EngineeringWeb CrawlingWeb ScrapingData PipelineCLI Development

ImmiCalm (Best Business Impact Award)

Winner of Best Business Impact at Nepal-US Hackathon 2026. Co-developed ImmiCalm, a privacy-first AI immigration assistant for Nepali F-1 and H-1B visa holders in the US, for Nepal-US Hackathon 2026. Built a multi-agent LangGraph system that intelligently routes questions between FAISS-based RAG retrieval, live web search, and Gemini response generation. Created personalized dashboards, severity-based news filtering, F-1/OPT timeline tracking, myth-busting workflows, and mood-aware guidance while keeping user profile data entirely client-side for privacy.

LangGraphAI AgentsRAGWeb SearchFastAPINextjsTailwindCSS

Automated Database Normalization Checker

A Python CLI tool that analyzes live PostgreSQL schemas using user-defined functional dependencies and automatically detects and decomposes relations up to Third Normal Form (3NF). Implements formal normalization algorithms inspired by database research, bridging relational theory with real-world schema design.

DatabasesRelational DatabasesPostgreSQLDatabase NormalizationSQLPython

Medical Chatbot using Multi-Agentic RAG

An AI-powered chatbot that provides medical information and preliminary diagnoses using advanced NLP techniques and medical knowledge bases.

Multi-Agentic RAGWeb SearchLLMLangGraphVector Database

Named Entity Recognition using BERT

Fine-tuned BERT model for Named Entity Recognition, capable of identifying and classifying named entities in text with high accuracy.

Named Entity RecognitionToken ClassificationTransformersBERTFine-tuningDeep LearningPytorch

IMDB Reviews Sentiment Analysis

Deep learning model for sentiment analysis of IMDB movie reviews, using transformers to classify reviews as positive or negative.

NLPSentiment AnalysisText ClassificationTransformersBERTFine-tuningDeep LearningPytorch

Finger Sign Recognition

Computer vision application that recognizes hand gestures and finger signs in real-time using deep learning models.

Computer VisionCNNClassificationDeep LearningSupervised LearningPytorch

MNIST classification using MLP from Scratch

A Multi-Layered Perceptron model that classifies MNIST dataset. The model uses Cross Entropy loss for evaluation loss and Stocastic Gradient Descent (SGD) for optimization. Also, ReLU and Softmax as activation functions.

MNIST ClassificationMLPSGDCross Entropy LossBackpropagationReLUSoftmax RegressionDeep Learning

Image Inpainting using GAN

A system that makes use of GAN to restore corrupted images and restore patches in the images.

GANPythonNodejs

Home Decor Marketplace with Recommendation System

A home decor marketplace built using MERN stack with a recommendation system implemented using Collaborative and Content-based Filtering.

Collaborative FilteringContent-based FilteringReactNodejsMongoDB