Project Overview
Walla Software partnered with Avahi to improve member retention for its fitness studio management platform. The goal was to transition from a prototype churn model to a production-ready solution powered by AWS SageMaker. Avahi implemented a scalable MLOps architecture including automated training, inference, and explainability pipelines. As a result, Walla now has real-time churn predictions with clear insights into why a member might drop out, helping studio owners act proactively.
About the Customer
Walla Software, Inc. is a SaaS provider based in San Diego, CA, serving boutique fitness studios specializing in yoga and pilates. Their platform offers scheduling, billing, staff coordination, and member engagement tools tailored for studio operators. Walla focuses on simplifying operations to drive client satisfaction and business growth.
The Problem
Walla needed a more accurate, scalable, and explainable solution for predicting gym member churn. The existing model was based on limited data and lacked integration into production systems. Without timely predictions and actionable insights, studio owners risked losing members without early warning. Walla wanted to improve model performance, automate retraining with new data, and make predictions interpretable for nontechnical stakeholders.
Why AWS
Walla selected AWS for its scalable machine learning capabilities, managed infrastructure, and integration flexibility. Amazon SageMaker offered a fully managed service to orchestrate training, retraining, and inference at scale. AWS also provided access to services like CloudWatch for monitoring, S3 for data storage, Lambda for orchestration, and RDS for result storage, supporting the end-to-end ML lifecycle.
Why Walla Chose Avahi
Avahi’s proven expertise in MLOps and AWS-native architectures made it the ideal partner for productionizing Walla’s churn prediction system. Avahi had previously delivered a successful proof-of-concept and was already familiar with Walla’s goals and data. Their deep knowledge of SageMaker pipelines, CI/CD automation, and realtime inference helped accelerate implementation and ensured a robust, maintainable solution.
Solution
Avahi redesigned Walla’s churn prediction platform by implementing a modular, automated MLOps pipeline on AWS:
- The solution ingests data from structured CSVs (e.g., bookings, subscriptions) via S3 triggers. A Lambda function initiates SageMaker pipelines when new data is uploaded.
- A preprocessing stage performs schema validation, feature extraction, and transformation.
- Avahi replaced the decision tree with a tuned XGBoost model, selected for its superior AUC performance and native feature importance capabilities. The model achieved an AUC of 0.846 on validation data.
- The SageMaker pipeline handles model training, evaluation, conditional registration, and deployment to endpoints.
- A monthly or performance-triggered retraining workflow ensures model freshness using SageMaker Pipelines.
- An inference pipeline provides real-time predictions with LIME-based explanations, helping users understand the key drivers behind each churn risk score.
- Avahi also implemented CI/CD for Lambda functions, managed in a dedicated GitHub repo with Docker-based builds and automatic updates.
Key Deliverables
- Automated ML pipeline for training, evaluation, registration, and deployment
- Real-time inference endpoint with LIME explainability
- Scheduled and performance-based retraining pipeline
- Deployment scripts, CI/CD integration, and Lambda automation
- Stakeholder-friendly explainability outputs
- Monitoring with CloudWatch and model versioning
Project Impact
The production grade MLOps system delivered by Avahi significantly enhanced Walla s ability to predict and act on gym member churn. Studio owners now benefit from automated insights that identify at risk members and the reasons behind their likelihood to churn, helping drive timely engagement and retention.
Metrics
- AUC score: 0.846 on validation set
- Model retrains monthly or on performance degradation
- Inference pipeline generates individual explanations with LIME
