An AI candidate ranking and discovery system built under strict constraints: CPU-only, 16GB RAM, 5-minute runtime. Features offline precomputation plus fast ranking using sentence-transformers, TF-IDF/BM25, and NumPy/Pandas scoring. Deployed as a Gradio app on Hugging Face Spaces.

A retrieval-augmented generation chatbot for HR policy Q&A, built with LangChain, FAISS, and Groq for a Kaggle-based RAG challenge, delivering accurate answer generation over complex documents.

Built for the FinCal Innovation Hackathon. A financial education and calculation platform with SIP, SWP, top-up SIP, goal-based planning, and retirement strategy modules designed to teach financial literacy through real-time, interactive visualization rather than static spreadsheets.

An embedding-based sentiment classification system for public posts on X, classifying text as Positive, Neutral, or Negative using semantic embeddings rather than keyword/TF-IDF matching, with XGBoost outperforming logistic regression on the embedding space.

A full semantic search and retrieval-augmented generation pipeline built on ~205,000 Flipkart product reviews. Dense embeddings via all-MiniLM-L6-v2, FAISS vector indexing for fast similarity search, and a RAG generation layer evaluated with Precision@K/Recall@K.

Built for the 2026 Meta PyTorch Hackathon. An OpenEnv-compliant reinforcement learning environment that trains AI agents to manage blood bank logistics, balancing emergency requests, strict blood-type compatibility, and FIFO expiry rotation across a graded, multi-difficulty evaluation system, deployable via Docker/Hugging Face Spaces.