Kavach
A support agent for Indian sellers that can ask for refunds but can't pay them. A separate checker re-reads the records first.
Built at AgentX Hackathon 2026
Team Paradigm: Anurag Rai, Abhishek Tiwari
Built with Python, FastAPI, LangGraph, PostgreSQL, Neon, Groq, Google Gemini, Twilio WhatsApp, Twilio Voice, Telegram Bot API, Gmail SMTP, Next.js, React, Tailwind CSS, PyTorch, Hugging Face Transformers, MuRIL, scikit-learn, Weights & Biases, Render, Vercel, GitHub Actions
About the project
Small online sellers lose hours to "where is my order?", return and refund messages. An AI agent can answer them, but an agent that can issue refunds can also be talked into the wrong one.
Kavach answers customers on WhatsApp, Telegram and the web in Hindi, Hinglish or English. It looks up orders and courier status and asks for refunds, but it cannot pay one. A separate verifier re-reads the real records and checks every fact the action depends on, including the exact policy line and the customer's photo (a vision check). It then allows, asks the seller (console, Telegram, email, or a phone call: "press 1 to approve"), or blocks. Only a payout service holding a signed one-time slip can write money, the AI's database user is refused writes by Postgres itself, and every step goes into a hash-chained audit log.
In our test of 19 attacks and 11 honest requests, Kavach paid ₹0 and blocked no honest customer. A plain strict-prompt AI paid on 6 of 54 attack tries. A guard model we trained for Indian chats (MuRIL teacher, distilled student, tracked in W&B) catches 73% of hand-written manipulation attempts, against 13% for rules and 7% for Meta's Prompt Guard 2. Every claim can be checked live on the Proof and "How it works" pages.