NCBI Gene MCP Client
Model Context Protocol (MCP) client integrating with NCBI Entrez API for gene and protein metadata retrieval.
- Symbol-based queries & flexible search with dual interfaces (REST API + CLI)
- Production-ready MCP server (JSON-RPC) compliant with NCBI rate-limiting
- Deployed on Vercel with modular design and automated testing
Skill Lyft: AI-Powered Learning Prototype
Learning platform simplifying complex ML concepts like Transformers, LLMs, and Neural Networks.
- Full-stack prototype with React frontend and Flask backend
- User testing with 20+ students in Agile 4-person workflow
- 30% improvement in navigation efficiency after usability testing
UniChatbot
LLM-powered chatbot using LangChain and RAG to answer 500+ university-related queries.
- Semantic search with vector embeddings, improving query relevance by 35%
- Optimized local inference with HuggingFace models and dynamic context windows
- Containerized with Docker and deployed via Streamlit
Smart StudyDesk: AI-Powered Study Organizer
Cross-platform study assistant that organizes lecture notes, summarizes content, and generates personalized quizzes.
- Usability testing with 20+ students using NASA-TLX workload analysis
- Improved navigation efficiency by 30% after iterative refinements
- Semi-structured interviews for user feedback
Zero-Shot Medical Image Classification with CLIP
Comparative study of CLIP against Vision Transformers (ViT) and ResNet50 for OCT medical image classification.
- Zero-shot classification with custom prompts
- Evaluated with Precision, Recall, F1, ROC-AUC, and MCC metrics
- Validated on OCT2017 and RFMiD datasets
HealthSenseAI: Automated Disease Detection
MERN stack web app with Hugging Face pre-trained model for automated disease predictions.
- End-to-end ML pipeline for preprocessing, inference, and deployment
- Leveraged Gradio to cut prototyping cycles by 40%
- User-friendly interface for non-technical healthcare staff
Efficient Model for Concrete Cracks Detection
Deep learning models for architectural infrastructure crack detection using knowledge distillation.
- Trained and benchmarked 9 CNN architectures on SDNET2018 dataset (56k+ images)
- Applied KD to compress ResNet152/DenseNet into lightweight models
- Improved MCC by up to 20% with Grad-CAM interpretability
Automated Face Recognition Attendance System
Real-time face recognition system using Haar cascade and OpenCV processing 200+ students per session.
- Deployed in Flask web app for remote access
- Reduced attendance processing time by 80%
Climate Change Forecast of Delhi
SARIMAX forecasting models on 20+ years of regional climate data to predict seasonal and long-term trends.
- Preprocessed climate time-series data (1.6B+ records, Berkeley Earth)
- Extracted Delhi-specific subsets for regional analysis
- Interpretable visual analytics showing warming trends