How ADEO Leroy Merlin decreased their time-to-market from 2 months to 2 weeks

• ADEO and it’s subsidiary Leroy Merlin, leading names in the retail sector, initiated a data-driven transformation journey.
• Adeo's time-to-market from idea to deployment, was hindered by manual processes, emphasized the need for a scalable ML pipeline framework.
• The challenges were evident: to streamline infrastructure complexities and enhancing team collaboration for ML model deployment, minimizing back-and-forth interactions between data scientists and ML engineers.
• They decided to abstract away these complexities with ZenML.
• Their time-to-market reduced from 8.5 weeks to a mere 2 weeks.
• They anticipate a remarkable 300% increase in deployment efficiency.

The Challenge: Manual Processes Burdening the ML Lifecycle

Traditional ML model development posed several challenges:

• Isolated Development
  Data scientists worked independently, facing issues with data versioning, model tracking, and conducting isolated experiments.
• Infrastructure Overhead
  Setting up computational resources for model training required DevOps support, posing difficulties for data scientists unfamiliar with Docker or specialized hardware.
• Reproducibility and Portability
  Switching between local and cloud environments for experiment reproducibility was complex.
• ML Infrastructure Complexity
  Managing the entire ML lifecycle across different tools and infrastructure increased workflow complexity.

The Solution: Abstracting away the complexity with ZenML

ADEO Leroy Merlin adopted ZenML to streamline ML pipelines without compromising flexibility. ZenML's framework facilitated easy pipeline construction, versioning, and deployment, emphasizing reproducibility and automation. Key features include:

• Quick Set-Up
  ZenML's Pythonic framework enabled swift pipeline annotation and construction.
• Data and Model Versioning
  Easy tracking and versioning of datasets and models ensured result reproducibility.
• Pipeline Portability
  ‍Abstracted infrastructure complexities allowed seamless prototyping and deployment across local and cloud environments.
• Framework and Infrastructure Agnosticism
  Flexibility to use any ML library, such as TensorFlow or PyTorch, and deploy workloads across different infrastructure targets with minimal overhead.

The Results: A Simplified and More Productive Workflow

Following a successful evaluation, ADEO Leroy Merlin conducted a pilot project using ZenML to classify penguin species based on physical traits, demonstrating its effectiveness. Subsequently,

• Centralized Experimentation Workflow
  ‍Data scientists utilized a centralized hub for logging experiments and sharing methodologies effortlessly.
• Autonomous ML Pipelines
  ZenML's abstraction layers empowered data scientists to independently set up pipelines for complex computational needs and deployment specifics.
• Enhanced Collaboration
  Predefined pipeline components available in ZenML enabled better reuse and sharing of common pipeline elements across teams.
• Accelerated Development Cycle
  Increased productivity allowed for rapid iteration, rigorous testing, and confident model deployment.
• Reduced Operational Overhead
  Automation of the ML lifecycle significantly decreased time spent on environment configuration and management, prioritizing high-value development tasks.

The Business Impact: A New Era of Retail Efficiency

Expedited Time-to-Market

The shift from 2 months to a mere 2 weeks from development to production marked a significant milestone in ADEO Leroy Merlin's ML journey.

Robust Deployment

With 5 ML models in production, the team is on track to reach a target of 20 by the end of 2024.

Automation with Operational Efficiency

1,614 development runs and 109 production runs showcased the reliability and confidence in the ML models deployed.

In order to roll-out a new business unit in another country, all the team needed to do was create one config file, and the model was already ready to go.

FTP Economy

Streamlined processes led to a remarkable reduction in FTP economy for the 6-member ML team.

Breaking Barriers

Data scientists and ML engineers expressed increased satisfaction as they autonomously deployed their models, shattering the 'stay in PoC' limitation.