█████╗ ██████╗ ██╗  ██╗██╗███╗   ██╗ █████╗ ███╗   ██╗██████╗  █████╗ ███╗   ██╗
██╔══██╗██╔══██╗██║  ██║██║████╗  ██║██╔══██╗████╗  ██║██╔══██╗██╔══██╗████╗  ██║
███████║██████╔╝███████║██║██╔██╗ ██║███████║██╔██╗ ██║██║  ██║███████║██╔██╗ ██║
██╔══██║██╔══██╗██╔══██║██║██║╚██╗██║██╔══██║██║╚██╗██║██║  ██║██╔══██║██║╚██╗██║
██║  ██║██████╔╝██║  ██║██║██║ ╚████║██║  ██║██║ ╚████║██████╔╝██║  ██║██║ ╚████║
╚═╝  ╚═╝╚═════╝ ╚═╝  ╚═╝╚═╝╚═╝  ╚═══╝╚═╝  ╚═╝╚═╝  ╚═══╝╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═══╝

abhinandan = {
    "pronouns": "he/him",
    "location": "Patiala, Punjab, India",
    "education": "B.E. ENC @ Thapar Institute of Engineering & Technology (2024–2028)",
    "current_focus": "Agentic AI systems, Backend Development",
    "communities": ["GDG on Campus — DSA Head"],
    "interests": ["Multi-Agent Systems", "RAG Pipelines", "Backend Systems", "LLM Engineering"],
    "actively_seeking": "ML/AI Engineering Internships @ AI Startups",
    "coding_style": "Build fast, evaluate rigorously, ship clean"
}

Calibrated payment dispute agent built for the Razorpay AI Buildathon 2026. A full agentic pipeline: an Ingestion Agent validates and encodes case data, a Prediction Layer runs an XGBoost classifier with Platt-scaled probability calibration plus SHAP-based explainability, and a Decision Agent branches into auto-fight / auto-drop / escalate-to-human, logging every decision to an audit trail. FastAPI backend with a Streamlit frontend; orchestrated end-to-end with LangGraph.

• Calibration improvement: Brier score 0.2322 → 0.2309, log loss 0.6564 → 0.6536
• Threshold-tuned via precision/recall and false-positive cost analysis

Multi-agent disaster prediction system (cyclone + earthquake) built during a research internship at NIT Delhi's Climate Lab. Combines cascaded XGBoost classifiers, ChromaDB RAG, and SHAP explainability, evaluated with RAGAS.

• Macro F1: 0.92 · Faithfulness (RAGAS): 0.797
• Targeting an IEEE INDICON submission

Multi-phone 3D motion capture fitness coaching system built at the Qualcomm Snapdragon Multiverse Hackathon. 4th place out of 60 teams. Led ML: camera calibration, DLT triangulation, One-Euro filtering, rep counting, and LLM-based coaching feedback.

• Reprojection error: 0.0767px · Jitter reduction: 41%
• LLM feedback via Gemini + Sarvam AI

Real-time payment fraud detection system combining Isolation Forest and Deep Autoencoders on 40,000+ government transactions.

🏆 Hack4Delhi National Finalist — Top 28 of 1,200+ teams