Revenue Intelligence

The Problem

Indian D2C founders selling on Amazon, Flipkart, and Shopify generate thousands of rows of sales data every month — but have no visibility into what's actually hurting their revenue.

Hiring a business analyst costs ₹50,000+/month. Most founders fly blind.

What We Built

Revenue Intelligence is a 4-agent AI pipeline that turns any sales CSV into a founder-ready action plan in minutes.

Upload a CSV → 4 AI agents chain together → get KPIs, insights, competitor research, and a Gemini-powered action plan.

How We Built It

Agent Pipeline

Agent 1 — Ingestion reads the uploaded CSV, auto-detects marketplace format (Amazon, Flipkart, Shopify, Meesho), cleans the data, and computes 18 business KPIs.

Agent 2 — Analysis runs 5 detection engines across the KPIs to surface revenue decline, weak states, fulfillment issues, and category underperformance.

Agent 3 — Research uses Gemini 2.5 Flash Lite with live Google Search grounding to research competitor pricing and market benchmarks for the top insights.

Agent 4 — Report reads all prior agent outputs and generates a prioritized founder action plan using Gemini.

MongoDB as Shared Agent Memory

MongoDB Atlas acts as the shared memory layer between all agents. Each agent reads the previous agent's output and writes its own — enabling full pipeline traceability across 6 collections.

The MongoDB MCP Server is registered in Google Cloud Agent Builder, allowing agents to inject real business context into every Gemini prompt.

Google Cloud Integration

Google Cloud Agent Platform — orchestrates the 4-agent workflow
Gemini 2.5 Flash Lite — powers research and report generation
Google Search Grounding — gives Agent 3 live web access for competitor research
MCP Server — registered at our live endpoint for context injection

Challenges We Faced

Managing memory limits on free-tier deployment while processing large sales CSVs with pandas
Handling Google Cloud ADC authentication on Render (solved with API key fallback)
Serializing pandas DataFrames to MongoDB and reconstructing them for Plotly rendering
Building a column mapping system that works across Amazon, Flipkart, Shopify, and Meesho CSV formats automatically

What We Learned

How to design multi-agent pipelines where MongoDB acts as shared memory between agents
How to use Google Search grounding with Gemini for real-time competitor research
How to register MCP servers in Google Cloud Agent Builder
The importance of graceful degradation — if one agent fails, the pipeline should still deliver partial value