Naoufal ACHARKI

Naoufal ACHARKI

Senior Data Scientist | Marketing ML Expert | Causal Inference Specialist

Marketing Campaign Recommendation System

Published:

Overview

Designed and deployed a sophisticated recommendation system to optimize customer targeting during digital marketing campaigns for a data infrastructure and analytics startup.

Challenge

Marketing teams struggled with:

  • Low conversion rates in digital campaigns
  • Inefficient customer targeting strategies
  • Lack of personalization at scale
  • No systematic approach to measure treatment effects

Solution

Technical Architecture

Built an end-to-end ML pipeline incorporating:

1. Data Engineering

  • Ingested and processed 50M+ customer interaction events
  • Built scalable ETL pipelines handling 3M customer profiles
  • Implemented feature engineering for behavioral signals

2. Machine Learning Models

  • Developed causal inference models to estimate treatment effects
  • Implemented uplift modeling to identify high-propensity customers
  • Used meta-learners (T-learner, S-learner, X-learner) for heterogeneous treatment effects

3. MLOps Infrastructure

  • Containerized models using Docker
  • Orchestrated workflows with Apache Airflow
  • Managed experiments and model versioning with MLflow
  • Deployed on AWS EC2 with Kubernetes for auto-scaling
  • Implemented CI/CD pipelines via GitLab

4. Production Deployment

  • Real-time scoring API using FastAPI
  • A/B testing framework for continuous optimization
  • Monitoring dashboards with Grafana

Technology Stack

  • Languages: Python, SQL
  • ML Libraries: Scikit-learn, XGBoost, CausalML
  • Data: AWS Redshift, PostgreSQL
  • Orchestration: Airflow, Kubernetes
  • MLOps: Docker, MLflow, DVC, GitLab CI/CD
  • Cloud: AWS (EC2, S3, Redshift)

Impact

  • 30% increase in conversion rates for targeted campaigns
  • 3 million customers scored in production
  • 50 million interaction events processed
  • Established scalable MLOps best practices across the organization
  • Reduced time-to-production for new models from weeks to days

Key Learnings

  1. Causal Inference Matters: Using uplift modeling vs traditional prediction models revealed significant improvements in identifying persuadable customers

  2. Production-First Mindset: Building with deployment in mind from day one avoided costly refactoring

  3. Experiment-Driven Culture: A/B testing framework enabled continuous iteration and measurable impact

Skills Demonstrated

  • End-to-end ML pipeline development
  • Causal inference and experimentation
  • MLOps and production engineering
  • Cloud infrastructure (AWS)
  • Cross-functional collaboration with marketing teams