I am a Full-Stack Developer & AI Engineer and an ICCR Scholar pursuing a B.Tech in Computer Science and Engineering at IIIT-Delhi. I specialize in bridging the gap between cutting-edge Artificial Intelligence (LLMs, RAG, Graph Neural Networks) and highly polished, modern web experiences.

Whether I'm architecting a real-time 3D dashboard in Next.js or training a machine learning model to autonomously detect security vulnerabilities, my goal is always to build robust, enterprise-grade software.

An advanced, AI-driven static application security testing (SAST) platform.

• AI/ML: Replaced legacy regex scanners with a hybrid GraphCodeBERT + GATv2Conv model that analyzes Code Property Graphs (CPGs) for structural vulnerabilities.
• RAG & RLHF: Integrated a Neo4j semantic knowledge graph (NVD CVEs) and fine-tuned StarCoder using Direct Preference Optimization (DPO) to autonomously generate compilable, security-correct patches.
• Full-Stack: Built the highly interactive, real-time 3D dashboard using Next.js 15, Framer Motion, and React Three Fiber. Wrapped the AI in a highly concurrent FastAPI backend.
• Integrations: Engineered a VS Code extension with real-time "Quick Fix" diagnostics and a GitHub Probot for automated PR security reviews.

A sleek, ultra-modern portfolio website showcasing my engineering and design capabilities.

• Built from the ground up using Next.js and Tailwind CSS.
• Features advanced UI/UX concepts including glassmorphism, scroll-driven animations, and interactive elements.

Autonomous Multi-Agent AI Platform for Scientific Research & Machine Learning Discovery.

• Architecture: Built an advanced multi-agent system using LangGraph, FastAPI, and Next.js.
• Capabilities: Features autonomous agents capable of orchestrating complex scientific workflows and data pipelines.

Enterprise-grade Retrieval-Augmented Generation (RAG) system built for production.

• Backend: Architected with PostgreSQL/PGVector for vector embeddings and semantic search, alongside Redis for high-speed caching.
• Infrastructure: Implemented SlowAPI for robust rate-limiting and deployed with fully automated CI/CD pipelines.

Production-ready crypto sentiment predictor and forecasting pipeline.

• AI/ML: Engineered an LSTM + Prophet ensemble model combined with FinBERT for advanced sentiment analysis and financial forecasting.
• Explainability: Integrated Explainable AI (SHAP) to provide transparent insights into model predictions, wrapped in a Streamlit dashboard.

An end-to-end, locally runnable Medical Language Model & Agent Framework.

• Model Architecture: Custom 350M parameter GPT-style Transformer built from scratch using Rotary Position Embeddings (RoPE) and SwiGLU activation.
• Training Pipeline: Pre-trained on 2GB of PubMed abstracts with memory-efficient techniques (Mixed-Precision fp16, Gradient Checkpointing) and SFT-tuned on medical dialogues.
• RAG & Serving: Integrated a FAISS vector database with ~49,000 embedded medical facts and wrapped in a OpenAI-compatible FastAPI backend with a custom web interface.

Explainable Machine Learning for early detection of atmospheric pollution trapping events in Delhi.

• Data Engineering: Unified 5+ years of CPCB air quality logs with global ERA5 meteorological reanalysis and NASA FIRMS fire hotspot data.
• Machine Learning: Engineered an XGBoost & HGBDT forecasting pipeline achieving 87.4% F1-score at 6h lead times, using meteorological deweathering to isolate atmospheric trapping events.
• Explainability: Applied SHAP interaction analysis to verify that model predictions align with physically correct boundary layer height interactions.

An autonomous, end-to-end multi-agent report generation and research system.

• Agent Network: Coordinated a state-machine network of 6 specialized agents (Supervisor, Researcher, Analyst, Writer, Reviewer, Output) using a dynamic routing architecture.
• Tech Stack & Streaming: Engineered with a hardened FastAPI backend and React v19 dashboard, using Server-Sent Events (SSE) to stream real-time node transitions.
• Human-in-the-Loop: Designed active HITL gate controls allowing human intervention before critical editing and compilation phases.