KPI Portal — AI-Driven Analytics Dashboard

Project Overview

How do you give executives a single-pane-of-glass view into millions of data points while keeping AI agents properly sandboxed? At Bajaj Finserv, I built the KPI Portal — a real-time analytics dashboard that consolidates performance metrics across the entire asset portfolio with sub-second query response for 2M+ concurrent users.

The twist? We needed to expose just enough data to AI agents for insight generation while keeping PII locked down tight.

The Challenge

Bajaj Finserv's data was scattered:

• Product metrics in one system
• User engagement in another
• OTP success rates somewhere else
• Campaign performance... you get the picture

Executives needed insights now, not after 3 days of data team wrangling. And the AI systems needed access — but not to everything.

Key Features

Dashboard Capabilities

• Multi-metric aggregation: DAU, Sessions, MAU, Bounce Rate — all in one view
• Real-time computation: Auto-refresh with WebSocket-powered live updates
• Dimensional filtering: Slice by product category, page type, traffic source, campaign
• Time-series analysis: 7-day, week-wise, month-wise, business-hours trends
• Comparative analysis: D18 vs D17 granularity for trend detection

AI Integration (The Interesting Part)

• Scoped Data Access: AI agents see only aggregated metrics — never raw PII
• Automated Alerting: Threshold-based alerts with executive notification
• Anomaly Detection: ML-powered insight derivation with recommended actions
• Complete Audit Trail: Every AI interaction logged for compliance

Performance at Scale

• Sub-second query response through optimized data structures
• 2M+ concurrent users with millisecond latency
• WebSocket real-time updates without polling overhead

Technical Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        CLIENT LAYER                             │
│    React.js + TypeScript + Tailwind + WebSocket Client          │
└────────────────────────────────┬────────────────────────────────┘
                                 │
                         HTTP / WebSocket
                                 │
┌────────────────────────────────┴────────────────────────────────┐
│                     APPLICATION LAYER                           │
│           Express.js REST API + WebSocket Server                │
│     JWT Auth │ Rate Limiting │ Data Aggregation │ Error Handling│
└────────────────────────────────┬────────────────────────────────┘
                                 │
         ┌───────────────────────┼───────────────────────┐
         │                       │                       │
    Cache Layer            Query Layer            AI Agent Layer
         │                       │                       │
    ┌────┴────┐          ┌──────┴──────┐         ┌──────┴──────┐
    │  Redis  │          │   Query     │         │  AI Service │
    │Sessions │          │   Engine    │         │  Insights   │
    │Metrics  │          │Aggregations │         │   Alerts    │
    │Alerts   │          │   Joins     │         │   Actions   │
    └────┬────┘          └──────┬──────┘         └──────┬──────┘
         └───────────────────────┼───────────────────────┘
                                 │
┌────────────────────────────────┴────────────────────────────────┐
│                         DATA LAYER                              │
│          MongoDB (Primary)  │  DocumentDB (Backup/Failover)     │
└─────────────────────────────────────────────────────────────────┘

AI Data Access Model (Security-First)

┌───────────────────────────────────┐
│     Full Data (MongoDB)           │
│  • Complete asset information     │
│  • User personal data             │
│  • Transaction history            │
└───────────────┬───────────────────┘
                │ Strict Filtering
                ↓
┌───────────────────────────────────┐
│    Scoped AI Agent Access         │
│  • Aggregated metrics only        │
│  • No PII exposure                │
│  • Time-windowed data             │
│  • Category-filtered views        │
└───────────────┬───────────────────┘
                │ Audit Logging
                ↓
┌───────────────────────────────────┐
│   Insight & Alert Generation      │
│  • Compliance verified            │
│  • Access timestamped             │
│  • Full audit trail               │
└───────────────────────────────────┘

Frontend Excellence

React + TypeScript Architecture

• Functional components with hooks for clean, testable code
• WeakMap caching for memory-efficient component data
• Tailwind CSS for responsive, accessible design
• Real-time WebSocket integration for live dashboard updates

Performance Optimizations

• Virtualized lists for large datasets
• Memoized selectors preventing unnecessary re-renders
• Lazy loading for non-critical dashboard sections
• Service worker caching for offline capability

Backend Scalability

Node.js + Express

• Async, event-driven for maximum throughput
• Connection pooling for database efficiency
• Request batching for API aggregation
• Comprehensive error handling with proper HTTP semantics

Docker Containerization

Docker Multi-container Deployment
├── API Service Container (Express.js)
├── Worker Service Container (Aggregation Jobs)
├── Redis Cache Container
└── Database Service Container (MongoDB + DocumentDB)

Results

Metric	Achievement
Query Response	<1 second
Concurrent Users	2M+ supported
AI Data Isolation	100% PII-free agent access
Dashboard Latency	Milliseconds
Uptime	99.9%+

Tech Stack

Frontend: React.js, TypeScript, Tailwind CSS, WebSocket Backend: Node.js, Express.js, Redis, MongoDB, DocumentDB Infrastructure: Docker, AWS Security: JWT, Role-based access control, Audit logging

Key Learnings

Building for scale taught me that caching strategy is everything. Redis wasn't just for sessions — it became our real-time metric store, alert buffer, and WebSocket pub/sub backbone.

The AI data isolation pattern we built became a template for the entire organization — proving that you can give AI powerful insights without compromising user privacy.

---

This project proved that enterprise dashboards don't have to be slow, ugly, or insecure. You can have all three: speed, design, and security.