Gale | Devpost

KPI Dashboard

Inspiration

Arca Continental serves thousands of small retail stores, and losing a customer often means losing a long-term source of revenue. The challenge is that churn is usually detected after the customer has already stopped purchasing.

We wanted to build a solution capable of identifying early warning signs of churn so that sales teams can take action before losing the customer. By transforming historical sales and operational data into actionable insights, we aimed to help Arca Continental prioritize retention efforts where they matter most.

What it does

Gale is an AI-powered churn prediction platform that helps ARCA Continental identify customers at risk of leaving.

Using machine learning, the platform calculates a churn probability for each customer and highlights the factors driving that prediction. Through an interactive Streamlit dashboard, users can:

Monitor key KPIs and monthly sales trends.
Identify customers with sales drops greater than 40%.
Filter customers by Customer ID, location, business channel, and number of detected risk signals.
Explore churn risk across territories and regions.
Receive AI-generated explanations and retention recommendations powered by Llama 3.2 through Ollama.

This allows sales teams to prioritize high-risk accounts and take action before revenue is lost.

How we built it

We worked with datasets containing approximately 200,000 customers and multiple business dimensions, including sales history, transaction activity, store size, cooler inventory, and cooler door counts.

After cleaning and consolidating the data, we performed extensive feature engineering to capture customer behavior over time. We created indicators related to sales decline, transaction slowdown, purchasing trends, and equipment utilization while removing redundant and highly correlated variables.

The final LightGBM model achieved:

98.3% Accuracy
81.2% Recall on churned customers
ROC-AUC above 0.98

To improve explainability, we integrated Llama 3.2 locally through Ollama, allowing users to understand why a customer is at risk and what actions could help retain them. The entire solution was deployed through a Streamlit web application.

Challenges we ran into

One of the biggest challenges was dealing with highly imbalanced data, as only 17.95% of customers churned. This meant that accuracy alone was not enough, and we had to focus on recall and meaningful business impact.

Another challenge was engineering features capable of capturing behavioral changes before churn while avoiding data leakage. We also needed to make model predictions understandable to business users by combining machine learning outputs with AI-generated explanations.

Accomplishments that we're proud of

Successfully processing and analyzing data from nearly 200,000 customers.
Achieving 98.3% accuracy and over 0.98 ROC-AUC.
Identifying that customer sales often decline dramatically during the six months preceding churn.
Discovering priority regions where retention efforts could have the greatest impact.
Delivering a complete end-to-end solution that combines machine learning, business intelligence, and generative AI.

What we learned

This project reinforced the importance of feature engineering in predictive analytics. We learned that understanding customer behavior is often more valuable than simply choosing a more complex model.

We also learned that explainability is essential for business adoption. Commercial teams need to understand not only which customers are at risk, but also why they are at risk and how they can respond.

What's next for Gale

Future versions of Gale will include real-time monitoring, automated retention alerts, CRM integrations, and personalized recommendations tailored to each customer profile.

We also plan to expand predictive capabilities beyond churn to include sales forecasting, territory optimization, and customer growth opportunities, helping businesses move from reactive decision-making to proactive strategy.