Featured Projects & Systems
Interactive catalog of end-to-end architectures: multi-agent autonomous loops, hybrid vector retrieval, medical computer vision, and machine learning pipelines.
System Architecture Flowcharts & Production Schemas
Every system is backed by verified GitHub Mermaid architecture diagrams (JWT/RBAC, RRF rank fusion, cross-encoders, and 4-agent state loops).
Enterprise RAG Platform
Built a production-oriented RAG platform for secure Q&A over internal company knowledge, enforcing JWT authentication and RBAC-based permission filtering. Designed a routing pipeline rewriting queries using conversation history, dispatching an Agent Planner across Document RAG (hybrid dense + sparse retrieval) and GraphRAG branches (bounded to 3 steps). Fused results with Reciprocal Rank Fusion (RRF) and cross-encoder reranking, enforcing evidence-before-answers grounded with Nemotron. Built a document ingestion pipeline with PostgreSQL & pgvector, exposed via FastAPI with Swagger docs.
Hybrid RAG Pipeline — Agentic RAG System
Ingested PDFs with pypdf, extracted and cleaned page-level text, and split into ~500-character chunks (50-character overlap) using RecursiveCharacterTextSplitter. Generated dense vector embeddings with HuggingFace all-MiniLM-L6-v2 (384-dim) and fit a BM25 sparse encoder into a Pinecone hybrid index. Implemented batched upserts and a PineconeHybridSearchRetriever (top-k=5) that feeds grounded context to Google Gemini 2.5 Flash (temp 0.2) with strict source file and page-level traceability.
Multi-Agent AI Research System
Constructed an autonomous 4-agent collaborative pipeline: Research Agent queries Tavily Search API for ranked web summaries, Reader Agent isolates clean content with BeautifulSoup, Writer Agent synthesizes structured reports with citations using Gemini 2.5 Flash, and Critic Agent audits reports with numerical scores (1–10) and feedback. Sequenced in a LangChain pipeline with Streamlit web UI (live logs, downloadable reports) and headless CLI mode.
Breast Cancer Classification using YOLOv11 & YOLOv26
Preprocessed and combined two public mammography datasets (CBIS-DDSM & VinDr-Mammo) into a unified classification dataset spanning benign and malignant cases. Applied 70:10:10 train/val/test splits, data augmentation (rotation, flipping, contrast adjustment), and class-balancing techniques. Ran systematic hyperparameter tuning independently on YOLOv11 and YOLOv26 classification heads, benchmarking precision, recall, and accuracy on held-out test data.
Object Detection using YOLOv8
Curated and labeled a custom dataset of cold-drink bottles with bounding-box annotations for supervised training. Applied data augmentation (flips, color jitter, mosaic) and transfer learning from pretrained COCO weights. Trained YOLOv8 to 92% detection accuracy, built an inference pipeline with configurable confidence and NMS thresholds, exported to ONNX format, and evaluated confusion matrices.
Predictive Analysis of Youth Suicide Risk
Designed a high-impact predictive classification system incorporating rigorous preprocessing, feature engineering, and SMOTE class balancing. Benchmarked Logistic Regression, Random Forest, and XGBoost (98.19% accuracy). Integrated SHAP explainability for interpretable feature importance and clinical risk factor transparency.
Hate Speech Detection Using Machine Learning
Collected and cleaned a labeled tweet dataset spanning hate speech, offensive language, and neutral categories. Preprocessed text (lowercasing, punctuation and stopword removal) and vectorized with CountVectorizer for bag-of-words. Trained a DecisionTreeClassifier for multi-class classification in an end-to-end pipeline.
End-to-End Data Analysis Pipeline
Built a comprehensive tabular data analysis pipeline in Python, Pandas, and NumPy to clean, transform, and explore large datasets. Generated automated visualizations and computed statistical summaries to identify trends, patterns, and actionable business insights.