churnshield

💡 Inspiration

Every day, millions of small sellers on Meesho and Flipkart silently lose customers — and they have no idea why.

I spoke to small sellers in Surat and found a common pain point: they could see customers disappearing, but had no tool to predict or prevent it. Enterprise CRM tools like Salesforce cost ₹15,000+/month — completely unaffordable for micro-sellers earning ₹2–8 lakh/month.

That gap inspired ChurnShield — an AI-powered churn predictor built specifically for India's 62M+ small e-commerce sellers.

How I Built It

ChurnShield uses a classic RFM (Recency, Frequency, Monetary) analysis pipeline combined with a machine learning model to score every customer's churn risk.

The pipeline:

1. Seller uploads order history as a CSV file
2. Python + Pandas cleans and engineers RFM features
3. A Random Forest classifier scores each customer:
   • 🔴 High Risk
   • 🟡 Medium Risk
     
   • 🟢 Loyal
4. Streamlit renders a simple visual dashboard
5. System generates a personalised WhatsApp retention message

RFM Score formula used:

$$ \text{RFM Score} = w_1 \cdot R + w_2 \cdot F + w_3 \cdot M $$

Where \( R \) = days since last order, \( F \) = order frequency, \( M \) = total spend, and weights \( w_1, w_2, w_3 \) are tuned on training data.

What I Learned

• How to translate a real business problem into an ML pipeline
• RFM analysis and its power for customer segmentation
• Building user-facing dashboards with Streamlit
• The importance of user research — talking to real sellers changed how I designed the product
• How to think about zero-cost deployment for early-stage products

Challenges I Faced

• No real dataset available — had to use a public Kaggle e-commerce dataset and adapt it to the Indian context
• Keeping it simple — the hardest part was resisting the urge to over-engineer and instead focusing on what a non-technical seller actually needs
• Solo project — managing research, design, ML, and presentation alone in limited time required strict prioritisation

Impact

| Target users | 62M+ small sellers in India | | Revenue recoverable | Up to 35% of churned revenue | | Cost to run MVP | ₹0 (Google Colab + Streamlit free tier) | | Time to build MVP | ~6 weeks solo |

What's Next

• Hindi UI for Tier 2/3 city sellers
• Direct API integration with Meesho & Flipkart
• WhatsApp Business API for automated retention messages
• Freemium model: Free for first 500 customers, ₹99/month beyond that

Built With

• google
• numpy
• pandas
• plotly
• python
• scikit-learn
• streamlit