InvoNotify — AI-Powered Invoice Revenue Recovery

Architecture & System Design →

An AI-powered invoice recovery platform that automatically detects overdue invoices, assesses risk using machine learning, and executes multi-channel recovery actions via Razorpay payment links.

Live Demo: invonotify.vercel.app Live Demo: YouTube

---

Table of Contents

1. Overview
2. Key Features
3. Quick Start
4. Architecture
5. Tech Stack
6. Environment Variables
7. Demo Setup
8. AI Recovery System
9. Dashboard Pages
10. API Reference
11. Razorpay Integration
12. Testing
13. Deployment
14. License

---

Overview

InvoNotify solves the ₹50,000 crore problem of unpaid B2B invoices. Instead of manual collection teams chasing payments, an AI agent autonomously:

1. Detects overdue invoices
2. Scores risk using ML (11 features)
3. Decides the best action and channel
4. Executes via email, WhatsApp, or Razorpay payment links
5. Learns from outcomes to improve

Demo Credentials:

• Email: razorpay@invo-notify.test
• Password: razorpay

---

Key Features

AI Recovery Engine

• ML risk scoring with 11 weighted features
• LLM-powered decision agent (GPT-4o-mini)
• 9 policy guardrails (no harassment, opt-outs, business hours)
• Explainable decisions with feature contributions

Multi-Channel Notifications

• Email — Gmail SMTP with branded templates
• WhatsApp — Meta Cloud API (free within 24h window)
• Razorpay Payment Links — One-click checkout

Customer Intelligence

• CIBIL credit score integration
• Payment history analysis
• Promise-to-pay tracking
• Automated reminder scheduling

Dashboard & Analytics

• Real-time recovery funnel
• Baseline vs AI comparison (85% vs 18%)
• Strategy effectiveness tracking
• Full audit trail

---

Quick Start

# Clone and install
git clone https://github.com/HEETMEHTA18/InvoNotify.git
cd InvoNotify
pnpm install

# Setup database
npx prisma generate
npx prisma migrate deploy

# Seed demo data
pnpm ai:seed

# Start dev server
pnpm dev

Open http://localhost:3000 and login with razorpay / razorpay.

---

Architecture

See proposed.md for full architecture diagrams and system design.

High-Level Flow

Overdue Invoice → ML Risk Scoring → Decision Agent → Policy Engine → Execute Action
                                                                     ↓
                                              ┌──────────────────────┤
                                              ↓                      ↓
                                           Email              WhatsApp
                                              ↓                      ↓
                                         Payment Link ←──────────────┘
                                              ↓
                                         Razorpay Checkout
                                              ↓
                                         Webhook → Case Closed

Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, Tailwind CSS
Backend	Next.js API Routes
Database	PostgreSQL (Neon) + Prisma ORM
Auth	NextAuth v5
Payments	Razorpay Test Mode
Email	Gmail SMTP (nodemailer)
WhatsApp	Meta Cloud API
ML	Logistic Regression (custom)
LLM	GPT-4o-mini
Hosting	Vercel

---

Environment Variables

Create .env in repository root:

Required

DATABASE_URL="postgresql://..."
DIRECT_URL="postgresql://..."
AUTH_SECRET="your-secret"
NEXTAUTH_SECRET="your-secret"

Razorpay

RAZORPAY_KEY_ID="rzp_test_..."
RAZORPAY_KEY_SECRET="..."
RAZORPAY_WEBHOOK_SECRET="..."

Email

GMAIL_USER="your-email@gmail.com"
GMAIL_APP_PASSWORD="your-app-password"

AI (Optional - uses fallback if not set)

OPENAI_API_KEY="sk-..."

Cron

CRON_SECRET="long-random-secret"

---

Demo Setup

Seed Demo Data

# Reset and seed 9 customers, 15 invoices
pnpm ai:seed

# Run AI sweep (safe demo mode)
curl -X POST http://localhost:3000/api/ai/recovery \
  -H "Content-Type: application/json" \
  -d '{"dryRun": true}'

Demo Dataset

Customer	CIBIL	Risk	Demonstrates
Acme Traders	720	LOW	Gentle reminder
Beta Industries	580	HIGH	Human approval required
Gamma Retail	810	LOW	Excellent payer
Delta Logistics	620	HIGH	Chronic late payer
Epsilon Foods	700	MEDIUM	Payment link
Zeta Pharma	690	BLOCKED	Opted out
Omega Constructions	750	VIP	Human handled
Kappa Textiles	660	STOP	Below cost floor
Sigma Electronics	540	CRITICAL	Escalate immediately

---

AI Recovery System

ML Risk Model

11 features scored via logistic regression:

{
  amountDue,           // 0.15 weight
  daysOverdue,         // 0.20 weight
  customerAgeDays,     // 0.05 weight
  previousInvoiceCount,// 0.08 weight
  previousLatePayments,// 0.18 weight
  averagePaymentDelayDays, // 0.12 weight
  paymentSuccessRate,  // 0.10 weight
  previousReminders,   // 0.05 weight
  isVipExempt,         // 0.02 weight
  cibilScore,          // 0.03 weight
  humanEngaged         // 0.02 weight
}

Risk Levels

Score	Level	Action
≥ 0.85	CRITICAL	Escalate to human
≥ 0.70	HIGH	Aggressive follow-up
≥ 0.40	MEDIUM	Payment link + reminders
< 0.40	LOW	Gentle nudge

Policy Guardrails

9 guardrails prevent:

• Contacting opted-out customers
• Harassment (frequency caps)
• Contacting outside business hours
• Auto-sending large amounts (requires approval)
• Chasing invoices below cost floor

---

Dashboard Pages

Route	Description
/dashboard	Main KPIs and metrics
/dashboard/invoices	Invoice list + create
/dashboard/recovery	AI recovery war room
/dashboard/recovery/analytics	Funnel + baseline comparison
/dashboard/diagnosis	ML failure classification
/dashboard/credit-scores	CIBIL score dashboard
/dashboard/promises	Promise-to-pay tracking
/dashboard/whatsapp	WhatsApp channel config
/dashboard/customers	Customer management
/dashboard/settings	Merchant settings

---

API Reference

Core Endpoints

• POST /api/invoices — Create invoice
• GET /api/invoices — List invoices
• POST /api/payments — Record payment
• POST /api/invoices/[id]/send — Send invoice email

AI Recovery

• GET /api/ai/recovery — List recovery cases
• POST /api/ai/recovery — Run AI sweep
• POST /api/ai/recovery/[id]/approve — Approve action

v1 API

• POST /api/v1/recovery-cases/[id]/diagnose — ML diagnosis
• POST /api/v1/recovery-cases/[id]/score — Risk scoring
• POST /api/v1/recovery-cases/[id]/promises — Create promise
• POST /api/v1/promises/reminders — Process reminders
• POST /api/v1/credit-score — Fetch credit scores

Webhooks

• GET/POST /api/webhooks/razorpay — Razorpay payment events

---

Razorpay Integration

Payment Link Flow

1. AI creates payment link via POST /payment_links
2. Link URL included in email + WhatsApp
3. Customer clicks → Razorpay checkout
4. Payment succeeds → Webhook fires
5. Invoice marked as paid → Case closed

Webhook Handling

// GET handler (callback_method: "get")
GET /api/webhooks/razorpay?razorpay_payment_id=...&razorpay_payment_link_status=paid

// POST handler (webhook events)
POST /api/webhooks/razorpay

---

Testing

# AI unit tests (15 tests)
pnpm ai:unit

# Full test suite (42 tests)
npx tsx --test tests/ai/agent/decision-agent.test.ts tests/ai/policy/engine.test.ts

# TypeScript check
npx tsc --noEmit

---

Deployment

Vercel

# Push to GitHub
git push origin main

# Vercel auto-deploys from main branch

Environment Variables (Vercel)

Set in Vercel Dashboard → Settings → Environment Variables:

• DATABASE_URL
• RAZORPAY_KEY_ID
• RAZORPAY_KEY_SECRET
• GMAIL_USER
• GMAIL_APP_PASSWORD
• CRON_SECRET

GitHub Actions

• CI: Runs tests on every push
• Recovery Sweep: Cron job every 6 hours

---

Project Structure

app/                    Next.js pages and API routes
├── dashboard/          Dashboard pages
│   ├── credit-scores/  CIBIL score dashboard
│   ├── diagnosis/      ML failure diagnosis
│   ├── invoices/       Invoice management
│   ├── promises/       Promise-to-pay
│   ├── recovery/       AI recovery + analytics
│   └── whatsapp/       WhatsApp channel
├── invoice/            Public invoice pages
└── api/                Backend API routes
    ├── ai/             AI recovery endpoints
    ├── v1/             v1 API endpoints
    └── webhooks/       Webhook handlers

lib/                    Shared utilities
├── ai/                 AI engine (ML, decisions, policies)
├── razorpay.ts         Razorpay integration
├── whatsapp.ts         WhatsApp Cloud API
├── credit-bureau.ts    Credit score fetching
└── mail-service.ts     Email + WhatsApp delivery

components/             React components
├── customers/          CIBIL meter gauge
├── recovery/           Recovery case components
└── ui/                 Base UI primitives

---

Documentation

• Architecture & System Design →
• PRD.md — Product requirements
• DEMO_SCRIPT.md — 5-minute demo script
• DEMO_VIDEO_PIPELINE.md — Video creation guide

---

License

Developed for Razorpay Hackathon. All rights reserved.