Chuyển đến nội dung chính

Lesson 24: Capstone — End-to-end ML project + demo

Complete a project according to the rubric: baseline -> pipeline -> tuning -> evaluation -> API -> monitoring -> 1-page report for business.

🧠 AI & ML — Lesson 23 Lesson 24: Capstone — End-to-end ML project + demo

Machine Learning: From Basics to Advanced

Part 4: Production, Explainability and Capstone

xdev.asia

Lesson objectives

  • Complete end-to-end ML projects according to clear rubrics
  • Present results from a technical and business perspective
  • Prepare portfolio ready for interview

Submission checklist

  • Problem description, metric and baseline
  • Preprocessing pipeline + reproducible training model
  • Validation and error analysis results
  • Meaningful explainability (SHAP/permutation)
  • The inferencing API is active and has run instructions
  • Monitoring plan (drift, retraining, alert)
  • 1-page report for stakeholders

Capstone scoring rubric (100 points)

  • Correct problem definition and metric: 10
  • Data quality and preprocessing: 15
  • Baseline and systematic improvement: 15
  • Model evaluation + CV + tuning: 20
  • Error analysis + result interpretation: 15
  • Serving + reset + run instructions: 15
  • Monitoring/retraining + business reporting: 10

Implementation instructions

  1. Choose a specific use case (churn, fraud, demand forecasting, pricing).
  2. Use a simple baseline before using a complex model.
  3. Only optimize the set metrics from the beginning, don't change the metrics midway.
  4. Pack all preprocessing into the pipeline to avoid leakage.
  5. Write a short README: how to train, eval, serve and monitor after deployment.

Expected output

You have a complete project to include in your CV/portfolio and demo during the interview.

Suggested capstone submission route

  1. Conclusion and measurable success criteria.
  2. Baseline pin, primary metric set and secondary metric set.
  3. Complete pipeline + error analysis report.
  4. Has a minimal inference demo (batch or API).
  5. Submit the final report according to the 100-point rubric.

Artifact is required

  • Dataset card describes data and risks.
  • Model card describes the scope of use, limitations and fairness notes.
  • Repo has instructions to rerun the entire process.

Self-assessment before submission

  • Are the results reproducible on another machine?
  • Has the risk of leakage or bias been stated transparently?
  • Is there a plan for monitoring if put into production?