InvoNotify

InvoNotify is an AI-powered invoice revenue recovery platform. Instead of sending the same late-payment email to everyone, it detects overdue invoices, assigns each one a collection-risk score, and runs the recovery through a policy engine before acting — with Razorpay payment links and webhook settlement completing the loop.

The system was designed end-to-end first (see proposed.md), then built out as a working platform: invoicing, customers, dashboard, and the AI recovery engine.

The Problem

Collections fail quietly:

• Late payments go unnoticed — nothing flags the invoice that just crossed its due date.
• Follow-up is risk-blind — a steady customer one week late gets treated the same as a repeat offender months overdue.
• There's no payment path — reminders without a way to pay inline turn a nudge into a chore.

How recovery works

Invoice created → due date passes → AI detects overdue → ML risk score
      → Policy engine (9 guardrails) → Email / WhatsApp → Razorpay link
      → Customer pays → Webhook → invoice settled, case closed

1. Detect — the AI engine sweeps for invoices past due.
2. Score — a logistic-regression model predicts collection risk on a 0–1 scale from eleven features: balance, days overdue, payment history, success rate, prior reminders, CIBIL score, VIP flag, and whether a human is already engaged.
3. Decide — an LLM decision agent proposes an action; the policy engine approves or blocks it.
4. Act — recovery runs through the channel the decision allows: a gentle email, a Razorpay payment link, or a WhatsApp reminder via the Meta Cloud API.
5. Settle — a Razorpay webhook records the payment, marks the invoice paid, and closes the recovery case.

ML risk scoring

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

The model is custom logistic regression trained on invoice and customer history — no closed platform, no black box. Every score is explainable back to its feature weights.

Policy engine guardrails

No recovery action fires unchecked. Every decision is gated by nine rules: skip already-paid invoices, respect opt-outs, hold disputes, route VIPs to humans, drop balances too small to recover, require approval above a money limit, enforce contact cooldowns, cap daily escalations, and only contact within business hours.

If an action is blocked, the case still gets a decision record — the audit trail keeps the full history for every invoice.

Architecture

Dashboard (Next.js) → AI Engine (ML + LLM) → Providers (Razorpay)
        │                    │                     │
        ▼                    ▼                     ▼
   PostgreSQL           Policy Engine          Channels
   (Neon + Prisma)      (9 guardrails)         (Email / WhatsApp)

• Next.js 16, React 19, Tailwind on the frontend; Next.js API Routes for the backend.
• PostgreSQL (Neon) with Prisma ORM — users, invoices, recovery cases, agent actions, promises-to-pay, and payment events are first-class models with real migrations.
• Razorpay payment links do the collecting; the webhook endpoint verifies signatures before recording payment.
• Email via Gmail SMTP (nodemailer), WhatsApp via the Meta Cloud API — the two recovery channels.
• NextAuth v5 for session auth, with user-scoped database queries.
• CI/CD — GitHub Actions workflows run CI, a scheduled AI recovery sweep, and automated reminders.

Security

Session-based auth, user-scoped queries everywhere, CRON_SECRET protected endpoints, Razorpay webhook signature verification, and a policy engine that blocks unauthorized actions — with an audit trail for every decision.

Roadmap

• WhatsApp channel rollout and configuration polish.
• Recovery analytics and ML failure diagnosis in the dashboard.
• Multi-merchant support and production hardening.