ML Engineering turns raw data into actionable insights. Cutting edge models studied in theoretical computer science are often under utilized in the industry. This is in part due to rightful skepticism regarding the model’s explainability and ability to generalize to new data outside of their training sets. In my experience, decision trees and regression are by far the most widely used models where safety is critical.

Sure, you can use LLMs to chat to your customers, but you should not use it to detect fraud just yet.

The step by step ML development workflow

  • Data Preparation
    • Detect and handle missing values
    • Detect and handle outliers
    • Feature engineering
      • scaling
      • encoding categorical data
      • feature extraction (my favorite is dimension reduction via Eigen values)
    • Test, train, validate split
  • Model Selection
    • Choose model based on task: classification, regression, clustering
    • Train multiple models
    • Validate models using cross-validation
  • Model Serialization
    • Save the trained model in a deployable format
  • Deployment
    • Ensure consistent dockerfile or python venv
    • Set up CI/CD to ensure latest model is always in production
  • Monitoring & Maintenance
    • Track usage and performance statistics
      • Prometheus & Grafana
    • Alerts: automated mail, pagerduty
    • Update the model
      • Use a task orchestrator such as airflow to retrain the model

Usecases

  • Customer analytics & segmentation
    • targeted ads
    • a/b testing
  • Churn prediction
    • predict the factors that lead customers to stop using your products or services
  • Demand forecasting
  • Fraud Detection
  • Price Optimization
  • Competitor analysis

While ML models are useful for automating routine tasks where large amounts of data is involved, executives often want more concise breakdowns of business metrics. That is where Business Analysts and Data Scientists come in.