payment-recovery-agent

PayResQ is an AI-powered payment recovery agent built for Razorpay's AI Buildathon (Track 03). It targets the problem of late authorization: when a UPI/card payment is disrupted mid-transaction, a customer's money can be debited without the merchant's system getting clean confirmation, leaving the payment in an ambiguous state for up to 3 days. Naive systems either retry too early, risking a double charge, or leave the customer without explanation.

The system detects ambiguous or failed payment states, reasons about the right response per case, executes it, notifies the customer, and logs every decision for a full audit trail. Its pipeline runs from a payment event through deterministic classification, LLM reasoning only for genuinely ambiguous cases, bounded gated execution (auto-retry, wait-and-reassure, notify, or escalate), and logging to a Supabase audit trail. The LLM never acts directly; it only proposes a diagnosis, and a separate deterministic layer decides whether to execute, with a double-execution guard.

A FastAPI backend receives real Razorpay webhook events (payment.failed, payment.captured), verifies their signature, and runs them through the full pipeline automatically. A React dashboard deployed on Vercel is live rather than a static mockup, and includes a "Simulate a live payment" panel that triggers a real event through the deployed backend so it can be watched being classified, reasoned about, and logged in real time.

A full batch run processed an 82-event batch of 80 simulated Razorpay-style payment events plus real Razorpay test-mode payments verified live via the API and webhook. The tech stack includes Python and FastAPI on Render, React (Vite) on Vercel, Razorpay test-mode API with HMAC-verified webhooks, an LLM via OpenRouter used selectively, Supabase (Postgres) for the audit log, and Recharts for charts.