Razorgrowth

RazorGrowth is a permissioned AI merchant-growth agent that is playable as a simulator. It was built for the Razorpay AI Buildathon 2026. In the simulator, a player runs a store for a simulated week while payments fail, and an AI agent proposes Razorpay recovery campaigns. The agent cannot spend money without approval, and a deterministic policy engine blocks unsafe proposals. The goal is to win back as much lost revenue as possible.

At its core is a tool-calling AI agent using Gemini function calling. It investigates a merchant's failed payments with five read-only tools, compares campaign options, and proposes one. The agent runs in a bounded loop, and its proposal must pass a strict schema, a deterministic policy engine, and explicit human approval. It can never spend money itself. The system also includes an evaluation harness that plays full simulated weeks with the agent, a heuristic, a tool-driven grid search, and no agent, on identical customers.

RazorGrowth provides safe payment execution with atomic approvals, deterministic Razorpay reference_ids with retry, backoff, and a safe halt, so a payment link is never created twice. It uses HMAC-verified, deduplicated webhooks and a sanitised audit trail. A deterministic market simulator includes a customer-behaviour model, an incremental-lift score, and a fair weekly leaderboard. The product includes Google sign-in with httpOnly session cookies, a React + TypeScript game UI with hand-built accessible SVG charts, Alembic migrations, 70 backend tests in CI, and deployment on Vercel and FastAPI Cloud.