SMB Sentinel AI

login page for specific business person to access their own dashboard
customers needing their attention and ai recommendation
CLTV, sentimental analysis and customer sales values
customer transaction data analysis
able to send email directly through dashboard
showcasing agents and their esecution rate
gemini reasoning of customer sentiment
mongodb mcp operations used

Inspiration

Small business owners in India — salon owners, cafe operators, electronics retailers — lose customers silently. A bad haircut, a cold delivery, an unresolved warranty complaint. By the time they notice, the customer is gone. We built SMB Sentinel AI to catch these signals in real-time and give business owners the power to act before it's too late.

What it does

SMB Sentinel AI is a multi-tenant customer intelligence platform where each business owner logs in and sees:

Real-time sentiment analysis of customer messages across WhatsApp, Google Reviews, Zomato, and Instagram
AI-generated recovery plans personalized for each unhappy customer
One-click action buttons to send recovery emails or WhatsApp messages directly
Root cause analysis showing WHY customers are unhappy (pricing, delays, quality)
Churn prediction scores identifying who's about to leave

The entire system runs on 6 AI agents that collaborate through MongoDB's Model Context Protocol (MCP) — with zero direct database calls.

How we built it

Architecture: A multi-agent pipeline where each agent has a specialized role:

Sentiment Agent — Analyzes customer messages using Gemini 2.5 Flash
Supervisor Agent — Routes to downstream agents based on severity
Churn Agent — Predicts churn probability using behavioral signals
Root Cause Agent — Identifies the operational failure category
Recovery Agent — Generates personalized recovery strategies
Executive Agent — Creates prioritized executive briefs

The MCP layer: Every MongoDB operation — saving workflows, storing agent findings, inter-agent messaging, task delegation — flows through the official MongoDB MCP Server via stdio transport. The agents don't know they're talking to MongoDB; they just call MCP tools like find, insert-many, and update-many.

The dashboard: A Streamlit multi-tenant app where 3 demo business owners (salon, cafe, electronics store) each have separate logins, separate customer data, and industry-specific insights.

Challenges we ran into

MCP response parsing — The official MongoDB MCP Server wraps results in <untrusted-user-data-UUID> security boundary tags. We had to build a regex-based extractor to pull JSON from these wrapped responses.
Gemini rate limits — With 6 agents processing 10 customers, that's 60+ Gemini calls. We implemented exponential backoff (2s → 4s → 8s → 16s) to gracefully handle 503/429 errors.
BSON ObjectId mismatch — MongoDB's _id field is a BSON ObjectId, but MCP passes it as a string. We couldn't use _id for task completion — had to switch to composite filters (workflow_id + agent_name).
Persistent MCP sessions — The MCP server runs as a subprocess. We built a background-thread architecture with an async queue to maintain a single persistent session across all agent calls.

Accomplishments that we're proud of

Zero PyMongo architecture — every single MongoDB operation (50+ per run) flows through the Model Context Protocol. We proved that MCP can completely replace traditional database drivers in a production-grade multi-agent system. The 6-agent pipeline processes 10 customers end-to-end with full inter-agent communication, shared memory, and autonomous actions — all through a single persistent MCP session. We also built a multi-tenant SaaS platform where different business owners log in and get completely personalized AI-powered insights with one-click recovery actions.

What we learned

MCP is a powerful abstraction layer — it completely decouples application logic from database specifics
Multi-agent systems need shared memory (not just message passing) to collaborate effectively
Small business owners don't want dashboards; they want action buttons that solve problems