Resume

• Designed and developed MiRAG, a multi-modal RAG QA system enabling context-aware answering across text, websites, PDFs and YouTube.
• Engineered modular chains using LangChain, integrating Google Gemini and embedding APIs with FAISS for semantic search and retrieval.
• Added chain routing and conditional fallback logic to switch between direct LLM QA and RAG based on input type and vectorstore status.
• Developed chunking, recursive text splitting, and transcript parsing modules for accurate retrieval from long and unstructured documents.
• Built dynamic chat memory using RunnableParallel for conversational history and context window optimization.
• Deployed a clean Streamlit interface with modular tabs for multi-source QA, PDF uploads, YouTube transcript processing, and chat history export.

• Collected and cleaned 1,100+ Kolkata property listings; engineered geospatial features with outlier handling.
• Trained and evaluated four supervised ML models: ElasticNet, Random Forest, Gradient Boosting, and XGBoost.
• Best result with XGBoost (RMSE = ₹30.30L, R² = 0.84) outperforming ElasticNet (RMSE = ₹36.35L, R² = 0.77).
• Deployed a Streamlit app with real-time inference, model selection, and feature contribution visualization.

• Implemented brute-force and KD-Tree-based k-NN in C++ for 734-dimensional Fashion-MNIST (10K samples).
• Achieved ~55× brute-force and ~2.5× KD-Tree speedups via OpenMP parallelization; maintained 84.04% accuracy.
• Designed a custom C++ vector library with element-wise ops and squared Euclidean distance computation.
• Built a CLI tool (via cxxopts) with runtime config for dataset path, neighbors, split ratio, algorithm, and OpenMP toggle.

• Built a custom CNN (~113K params) for 18 fruit classes using ~23.6K images with stratified splits and normalization.
• Achieved 88.9% test accuracy (F1 ≈ 0.82) using conv blocks, batch norm, dropout, and global average pooling.
• Fine-tuned MobileNetV2 (ImageNet) with last 30 layers unfrozen; used mixed precision and Dropout (0.4).
• Achieved 99.35% test accuracy with early stopping, model checkpointing, and dynamic class extraction.

• Implemented a URL shortener with FastAPI backend and MySQL database for end-to-end URL management.
• Designed Base62 encoding for auto-incremented IDs to generate unique short URLs with click tracking.
• Developed API endpoints for URL shortening, redirection, and health checks with input validation and error handling.
• Built a minimal web interface with HTML/CSS and Flask, integrating frontend form submission with the backend API.

C, C++, Python, SQL, HTML, CSS, Streamlit

C++ STL, Pandas, NumPy, Scikit-Learn, TensorFlow, Keras, XGBoost, LangChain, FAISS, Hugging Face, Flask, FastAPI, Matplotlib, Seaborn

Supervised Learning, Regularization, CNNs, k-NN, Transfer Learning, RAG Pipelines, Semantic Search, Model Evaluation

Git, GitHub, MySQL, CMake, OpenMP, VS Code, Jupyter Notebook, Google Colab, Streamlit Cloud

Linux, Windows

Data Structures & Algorithms, Competitive Programming, Technical Communication

Data Structures and Algorithms • Object-Oriented Programming • File Organization and Database Systems • Operating Systems Design • Graph Theory and Algorithms • Artificial Intelligence and Machine Learning

• Mentor 80+ students in core DSA concepts through weekly tutorials, coding sessions, and algorithm debugging.
• Evaluate lab submissions, conduct viva assessments, and deliver structured feedback across all components.

Collected and conveyed student feedback, suggestions, and issues to the Central Library administration.

Awarded for excellence in short story writing and creative contributions on the platform.

Creative writing, digital art, mobile photography, literature and storytelling.