You're building a churn prediction system for a mobile network operator in Lagos -- a model that predicts which subscribers will leave, served as an API the retention team can query weekly.
The discipline skills: data profiling, preprocessing with documented decisions, training a classifier on imbalanced data, evaluating with per-class metrics, and serving a model as a FastAPI endpoint.
The AI-direction lesson: this is your first time directing AI through a full ML workflow. AI will produce code that runs, models that train, and metrics that look good. The gap between "it runs" and "it's correct" is the entire project. AI defaults to accuracy on imbalanced data -- 92% accuracy, near-zero recall on the class that matters. It preprocesses without explaining assumptions. It generates endpoints that handle the happy path. Everything you need to catch these failures is provided: the brief, the dataset, the evaluation criteria. Your job is to check AI's output against those targets and notice when "looks right" and "passes verification" are different things.
Your Role
You're building Emeka's churn prediction system. You direct Claude Code through the entire workflow — from profiling the data to serving a live API endpoint. AI writes the code. You provide direction, verify the results, and communicate back to Emeka.
This is the first real project. Everything you need is provided: the brief, the dataset, the algorithm, the evaluation criteria, the ticket breakdown. Your job is not to decide what to build. Your job is to direct AI accurately and verify that what it produces actually works for Emeka's problem.
What's New
Setup gave you the tools. This is where you use them for real work.
What changes is the shape of the work itself. You'll move through a full ML workflow — data profiling, preprocessing, training, evaluation, serving — directing AI through each stage. Every stage produces something concrete you can check against specific criteria.
Tools
- Python / pandas — data loading and profiling
- scikit-learn — preprocessing, training, evaluation
- Jupyter — notebook workflow
- FastAPI / uvicorn — model serving
- Claude Code — AI direction
- Git / GitHub — version control
- curl — API testing
Materials
Everything is provided for this project:
- Emeka's email with the full brief
- A clean subscriber dataset
- A data dictionary describing every column
- Algorithm and evaluation criteria (RandomForest, recall ≥ 0.55 on the churn class)
- A complete ticket breakdown for every stage of the work
You won't need to decide what to build or how to structure the work. The materials handle that. Focus on directing AI through each ticket and verifying what it produces.