About the Client
ETP Group is a global retail technology company with over 35 years of experience in empowering retailers to succeed in a rapidly evolving market. Over time, it has helped more than 500 brands across 17+ countries connect their in-store, online, and mobile operations through unified commerce solutions. As a result, retailers can streamline processes, enhance customer experiences, and drive sustainable growth in an increasingly digital world.
Business Situation and Requirements
The client works in a multi-tenant AI/ML environment and recognized that their infrastructure needed an upgrade. They wanted smoother model training, standardized deployments, and automated onboarding to support scalable growth, stronger governance, and consistent service delivery across all customer environments.
To address these challenges, the client partnered with Unthinkable to modernize and automate the training and deployment of its AI/ML models. Their goal was to simplify model training and deployment across tenants, reduce manual work, and ensure scalable, reliable, and efficient operations.
Their key requirements were:
Modernize and automate the existing AI/ML infrastructure to eliminate manual dependencies and improve overall operational efficiency.
Build a streamlined framework to automate model training and deployment across multiple tenants, ensuring speed, consistency, and reliability.
Implement standardized deployment processes to maintain uniformity and reduce operational disparities across environments.
Develop a scalable architecture that removes deployment bottlenecks and supports seamless business growth.
Accelerate tenant onboarding through automated environment provisioning and configuration tailored to diverse organizational needs.
Enhance security, compliance, and governance with automated controls, centralized monitoring, and optimized model lifecycle management.
Solution
Our team built a production-ready MLOps system to support scalable anomaly detection and recommendation engines in a multi-tenant setup. We used MLflow, Apache Airflow, CI/CD pipelines, and cloud-based infrastructure to make it reliable and efficient.
Moreover, we focused on automating the training and deployment of AI/ML models, turning experimental models into secure, real-time, tenant-specific deployments. This also removed manual work and automated onboarding. As a result, the system can run frequent predictions while maintaining scalability, strong governance, and smooth operations.
We applied this automation to two key use cases: anomaly detection and the recommendation system.
Anomaly Detection Model
- Robust Anomaly Detection Model: Implemented a model capable of identifying outliers across complex datasets, starting with detailed notebooks converted into production-ready code.
- MLflow tracked experiments, metrics, and data transformations for reproducibility and stakeholder visibility.
- Automated Model Training Workflow: Built a system to detect models needing retraining from database inputs. Apache Airflow schedules weekly runs, keeping anomaly detection models up to date automatically.
- Automated Deployment & Real-Time Predictions: Models are auto-deployed using Airflow DAGs. Batch anomaly predictions run every 20 minutes, with results stored in GCP buckets for easy access.
- Dynamic Tenant Onboarding: New tenants automatically have models trained and activated. MLflow tracks performance and provides clear prediction insights.
- End-to-End Automation & Efficiency: Tenant configuration is fully automated. Cleanup scripts remove redundant data, ensuring scalability and efficiency for high-volume workloads.
Recommendation System Automation
- Automated, Scalable Recommendation Engine: Ensures recommendations remain accurate and up to date as the user base grows, without manual intervention.
- User-Centric Item Recommendations: Analyzes user behavior and preferences to provide highly personalized suggestions, boosting engagement, conversions, and retention.
- Item-Centric User Targeting: Identifies users most likely to engage with a specific item, helping promote relevant, new, or trending content effectively.
Item-to-Item - Recommendations: Suggests related or complementary items based on co-occurrence patterns, encouraging discovery, cross-selling, and deeper engagement on the platform.
Impact
With the successful automation of the Anomaly Detection and Recommendation System, ETP significantly improved its operational efficiency. By replacing manual processes with advanced automation, the team not only reduced errors but also made tenant onboarding faster and smoother.
Furthermore, encouraged by these results, ETP entrusted Unthinkable to automate the Forecast model as well. This ongoing collaboration drives continuous innovation, strengthens the retail software ecosystem, and ensures further improvements in efficiency, accuracy, and overall business performance.
500+
Enterprise Software Projects
300+
Brands
35,000+
Stores
