Arca Guard

Dashboard View
Segmentation View
Territories View
AI Plan View
AI Agent View
Trends View
Clients View

Inspiration

The challenge only required delivering an Excel file containing churn predictions generated by a machine learning model. However, we believed that predicting churn alone was not enough. Businesses need actionable insights and tools that help employees respond effectively to at-risk customers.

That is why we decided to build a complete solution instead of a standalone model. We created a platform that not only predicts which customers are likely to churn, but also helps users understand the results through interactive dashboards, receive AI-generated retention strategies, and practice customer conversations using voice AI. Our goal was to bridge the gap between prediction and action.

What it does

Arca Guard predicts which customers are most likely to churn and provides tools to help commercial teams take action before customers are lost.

At its core, the solution uses a HistGradientBoostingClassifier built with scikit-learn. The model analyzes historical purchasing behavior and generates a churn probability score for each customer.

Key highlights:

Binary classification model that predicts whether a customer is likely to churn in the following month.
Temporal validation to avoid data leakage.
Handles highly imbalanced data (~0.8% churn rate).
Uses customer behavior patterns rather than demographic information.
Achieves a PR-AUC of 0.327, outperforming the logistic regression baseline by approximately 54%.
Captures 87% of future churners by targeting only the top 10% highest-risk customers.

The model generates a downloadable Excel file containing customer predictions and risk scores. Beyond that, our web platform provides:

Interactive dashboards for risk analysis and segmentation.
Territory-level and customer-level insights.
AI-generated retention recommendations powered by Gemini.
A voice-based coaching assistant powered by ElevenLabs that helps employees practice customer retention conversations and improve communication skills.

This transforms raw predictions into actionable business decisions.

How we built it

We built an end-to-end churn prevention platform.

The process begins with historical monthly sales data, from which we generate 32 time-based features without data leakage. These features are used to train a HistGradientBoosting model with temporal validation, producing churn probabilities for each customer.

The results are then visualized in a React dashboard that allows users to explore customer risk, identify trends, and prioritize retention efforts.

To increase the practical value of the predictions, we integrated generative AI technologies:

Gemini analyzes model outputs and generates personalized retention plans and recommendations.
ElevenLabs provides a conversational voice coach that helps employees prepare for customer interactions and practice handling objections.

Components delivered:

Data pipeline: Feature engineering and temporal dataset generation.
Predictive model: HistGradientBoosting with temporal validation and competition submission generation.
Insights layer: Aggregated analysis by territory, cooler availability, customer size, and time.
Web dashboard: Multiple interactive views powered by React and TypeScript.
AI activation layer: Gemini for retention strategies and ElevenLabs for conversational coaching.

Tech Stack:

Machine Learning: Python, pandas, scikit-learn, Parquet.
Frontend: React 19, TypeScript, Vite, Tailwind CSS, Recharts, react-simple-maps.
Generative AI: Google Gemini API and ElevenLabs Conversational AI.

Arca Guard is not just a churn prediction model—it is a system that predicts, prioritizes, and activates action.

Challenges we ran into

One of our biggest challenges was developing a machine learning solution despite none of our team members specializing in data science. We had to learn key machine learning concepts, experiment with different approaches, and understand how to validate predictive models correctly.

Another challenge was connecting predictive analytics with business action. Building an accurate model was only part of the problem; we also needed to create meaningful tools that would help users understand and act on the predictions.

Accomplishments that we're proud of

We are proud that our solution goes far beyond generating a prediction file.

We successfully developed a machine learning model from scratch, created an interactive analytics platform, integrated generative AI for strategic recommendations, and implemented a voice-based coaching assistant using ElevenLabs.

Most importantly, we transformed a predictive model into a practical business tool that helps users make informed decisions and improve customer retention.

What we learned

Throughout the project, we learned how business rules and domain knowledge can be combined with machine learning techniques to solve real-world problems.

We gained hands-on experience in feature engineering, model validation, churn prediction, data visualization, and generative AI integration. We also learned the importance of translating technical outputs into actionable business insights.

What's next for Arca Guard

Our next step is to move from mock data to a fully connected production environment.

We plan to:

Connect the platform to a live database instead of static datasets.
Automate prediction generation and updates.
Enable real-time customer monitoring.
Expand AI-generated recommendations with more personalized strategies.
Integrate directly with customer relationship management (CRM) systems.

Our vision is to make Arca Guard a complete customer retention platform that helps businesses identify risks early and take meaningful action before customers churn.