This project implements a sophisticated multimodal RAG system transforming YouTube videos into queryable knowledge bases through advanced frame extraction and caption analysis. Leveraging Gemini for inference and Qdrant for efficient vector storage, the system processes both visual and textual content to generate precise, timestamped responses to natural language queries.

Architected a multimodal WhatsApp memory assistance pipeline (FastAPI + AsyncIO) that ingests text/voice/images via Twilio webhooks, classifies intent in real time via Groq LLM inference, embeds memories in Mem0's vector store for semantic recall, and serves analytics through idempotent SQLite transactions.

Developed a full-stack AI-powered resume analyzer using React, TypeScript, and Claude Sonnet integration, featuring real-time PDF processing, multi-dimensional scoring system (ATS, content, structure, skills), and comprehensive feedback generation for job seekers, integrating Zustand state management, Tailwind CSS, React Router, and Puter.js services for authentication, file system operations, and data persistence, delivering a responsive user interface with drag-and-drop functionality and visual score components.

Developed a sophisticated tutoring chatbot leveraging Google's Agent Development Kit (ADK) principles with intelligent orchestration between specialized Math and Physics agents powered by Gemini API, integrating context-aware conversation management, autonomous query classification pipeline routing student queries to domain-specific agents and provide personalized responses through prompt engineering and tool integration.

Coding assistant built with LangGraph, simulating a multi-agent developer workflow to generate complete projects from natural language prompts. It utilizes Planner, Architect, and Coder agents to sequentially design, structure, and implement applications, leveraging tools for file I/O and code execution. The system is deployed with a FastAPI backend and a NiceGUI frontend for user interaction and project management

Implemented a database storing string key-value pairs using Raft Consensus Algorithm, ensuring consistent data replication and fault recovery across the distributed network of nodes and utilized the leader lease mechanism, similar to those used by geo distributed databases such as Cockroach DB and YugaByte DB.

Developed MetroSense, a novel web-based platform to empower visually impaired individuals navigate the Delhi Metro system, achieving 65.1% mAP@50 for identifying environmental elements from real-time image captures. Integrated LLAMA Vision 3.2 90B for sophisticated VQA, engineered with context-rich, few-shot prompting and optimized decoding parameters to achieve a BERT F1 score of 0.85, delivering semantically accurate, context-aware voice-synthesized responses to user queries for improved safety and autonomy.

Developed and benchmarked novel models (GAN, Neural ODE VAE, Neural ODE Classifier) for stock market analysis, achieving a 15% F1 improvement and 85% faster training via Neural ODEs.Implemented a CNN-LSTM architecture delivering high-accuracy regression (R² 0.99, MAE 143.58 on S&P 500) across five major indices on the CNNPred dataset.

Created a distributed online marketplace system, architected to facilitate direct transactions between buyers and sellers through a central platform hosted on Google Cloud VM instances, leveraging gRPC for communication and Protocol Buffers for efficient data serialization.

Implemented a distributed Map-Reduce framework comprising of Master, Mapper and Reducer components to perform K Means Clustering on a given dataset ensuring fault tolerance for both components and utlized gRPC for communication among the three processes for each iteration.

Implemented a simulator for a processor based on RV32I variant of RISC-V ISA where the microarchitecture included a 5 stage pipeline allowing forwarding/bypassing and separate execution unit for Network on Chip operations,along with a 2-way set associative cache following Least Recently Used replacement policy.