Sentinel — AI Revenue Recovery

Razorpay Hackathon · Track 03 — AI Revenue Recovery
Catch revenue before it's gone. Win it back — automatically, and safely.

Live demo · Console · API docs · Architecture · Evaluation

---

Try it in 30 seconds: open the live demo → Open console → Run recovery on batch. (The API is on a free tier — the first request after it's been idle takes ~30–50s to wake, then it's instant.)

Sentinel is an AI agent that recovers revenue lost to failed payments — it detects a failure, diagnoses the real reason, chooses one bounded recovery action, and executes it, with a deterministic policy engine enforcing every boundary and an immutable audit trail recording every step.

Results

Metric	Result
Recovery Rate	66.4%
Money Recovered	₹3.75L
Improvement vs Baseline	+₹90,505
Fraud Cases Safely Blocked	100%
Wasted Retries Prevented	123
Cases Evaluated	67

Measured against a naive retry-everything baseline.

On reproducibility: these figures come from one seeded batch and shift on each re-seed (the batch is regenerated). Reproduce any time with Re-seed → Run recovery on batch, then read /api/metrics. The live demo may show slightly different numbers than the table above for exactly this reason — baseline and Sentinel are always computed from the same constants, so the comparison stays fair.

The live console — money recovered, recovery rate vs baseline, net benefit, and every recovery it made.

---

The Problem

Indian subscription businesses lose revenue every day due to involuntary churn:

• Insufficient funds
• Expired cards
• Gateway failures
• RBI e-mandate authentication failures
• Payment risk blocks

Most systems blindly retry failed payments. Blind retries waste money, annoy customers, and still fail to recover revenue.

---

Real Problem Solved — RBI e-Mandate Failures

₹15,000+ Mandate Authentication Failures

When recurring payments require additional customer authentication, retrying cannot solve the problem. Sentinel detects these cases, classifies them correctly, and routes them into a mandate recovery workflow instead of wasting retries.

Metric	Result
Cases Detected	12
Cases Recovered	11
Revenue Recovered	₹2,98,532
Recovery Rate	91.7%

Highest-performing recovery workflow in the batch.

---

From Failed Payment to Recovered Revenue

Four steps. One intelligent recovery loop. No manual chasing. No blind retries.

flowchart LR
  D[Detect] --> Di[Diagnose] --> De[Decide] --> R[Recover]
  R -.-> D

Loading

Recovery Decision

Each failure is diagnosed into a class, mapped to one bounded action, and only executed if the policy engine allows it.

flowchart LR
  A[Failed Payment] --> B[Diagnose Failure]

  B --> C1[Insufficient Funds]
  B --> C2[Card Expired]
  B --> C3[Gateway Timeout]
  B --> C4[Fraud Risk]

  C1 --> D1[Delayed Retry]
  C2 --> D2[Update Card Link]
  C3 --> D3[Smart Retry]
  C4 --> D4[Block and Escalate]

  D1 --> E[Policy Engine]
  D2 --> E
  D3 --> E
  D4 --> E

  E --> F[Recovered]
  E --> G[Stopped]

Loading

---

How Sentinel Works

flowchart TD
  A[Detect] --> B[Diagnose] --> C[Decide] --> P[Policy Check] --> E[Execute] --> V[Verify and Audit]

Loading

• The AI diagnoses.
• The policy engine decides what is allowed.
• The tools execute.
• Every action is audited.

---

System Architecture

Sentinel separates diagnosis from execution. The LLM decides what happened; the policy engine decides what is allowed; tools execute only approved actions.

Components

• React + Vite Frontend
• Node.js + Express Backend
• SQLite + Prisma
• Claude → Gemini → Rules Fallback
• Razorpay Integration (test mode — real test-mode payment links)
• Twilio Voice (test-mode / simulated when keys absent)
• WhatsApp Recovery (test-mode / simulated when keys absent)
• Audit Layer

---

AI Agent Design

Diagnosis Engine

Claude → Gemini → Rules. Returns strict JSON:

{
  "class": "...",
  "confidence": 0.92,
  "why": "...",
  "action": "..."
}

Policy Engine

• Max 3 retries
• Max 3 contacts
• Never contact fraud
• Never touch paid cases
• TRAI calling window

Tool Executor

12 recovery tools.

---

Evaluation

Measured, Not Marketed

Metric	Sentinel	Baseline
Recovery Rate	66.4%	50.4%
Money Recovered	₹3.75L	₹2.84L
Extra Revenue	+₹90,505	—
Retries Fired	38	161
Fraud Contacts	0	N/A

Dataset

• 67 failed payments
• Razorpay error schema
• 5 failure classes
• Realistic INR amounts
• Zero PII

---

Example Recovery Story

₹48,000 Mandate Failure

Failure	mandate_afa_required
Diagnosis	Mandate Re-Authentication Required
Action	represent_mandate
Result	Recovered
Revenue	₹48,000

---

Safety & Compliance

• Deterministic guardrails
• RBI-aware recovery flows
• TRAI call windows
• Fraud auto-block
• Idempotent execution
• Immutable audit trail

---

Live Recovery Demo

Sentinel can generate a real Razorpay test payment link. When the payment is completed:

flowchart LR
  F[Failed] --> L[Recovery Link] --> P[Payment Completed] --> W[Webhook Received] --> R[Recovered]

Loading

Dashboard metrics update automatically.

See it for yourself — a real link Sentinel generated

This is an actual payment link Sentinel created by calling Razorpay's Payment Links API for a failed ₹32,000 charge (customer Vivaan Rao):

💳 Click here to open the real Razorpay test checkout →

It opens Razorpay's test-mode hosted checkout (razorpay.com/payment-link/…/test) — this is exactly what the customer receives. No real money moves. Pay it with any Razorpay test card (e.g. 4111 1111 1111 1111, any future expiry, any CVV) and — with the webhook configured — the case flips to Recovered on its own.

Under the hood, the agent's create_payment_link tool returned:

{
  "tool": "create_payment_link",
  "result": {
    "url": "https://rzp.io/rzp/8wQFk0YK",
    "id": "plink_TWo99whbvLuZkB",
    "live": true
  }
}

"live": true means it's a real Razorpay test-mode link (not a simulated placeholder).

---

Repository Structure

SENTINEL/
├── client/       # React + Vite console (landing + dashboard)
├── server/       # Node + Express API, agent, tests
│   ├── prisma/   # SQLite schema
│   └── src/      # agent, config, seed, index, tests
└── *.md          # QUICKSTART · DEPLOY · ARCHITECTURE · API · EVALUATION

---

Quick Start

git clone https://github.com/Manvi0408/Sentinel.git
cd Sentinel
npm install
npm run setup
npm run seed
npm run dev
# open http://localhost:4100

Runs with no keys on the built-in rules engine. Add TEST-mode keys in server/.env (copy from server/.env.example) to enable LLM diagnosis and real Razorpay test links.

---

Documentation

• QUICKSTART.md
• DEPLOY.md — split backend/frontend deploy (Render + Vercel)
• ARCHITECTURE.md
• API.md
• EVALUATION.md

---

Key Insight

The reliability of an AI agent comes from what it is not allowed to do.

Sentinel puts the LLM behind a deterministic policy engine, making autonomous recovery measurable, auditable, and safe.