Same Nairobi veterinary clinic, new question. This time you predict which appointments are likely to be no-shows so the team can send extra reminders or double-book the risky slots.
The discipline skills: preparing data for regression with scikit-learn, building and evaluating a prediction model, comparing model performance against a naive baseline, and translating model accuracy into language the client can use.
The AI-direction lesson: you write your own prompts this time, and the size of each request matters. "Do the entire analysis" produces worse results than focused, sequential steps. AI also has a default that will inflate your model's accuracy -- random train/test splitting on data with a time dimension. The verification targets will help you catch it, but the skill is learning to question AI's structural choices, not just its numerical output.
Your Role
You direct AI through a structured prediction pipeline: prepare the data, build the model, evaluate it honestly, and translate the results into something the clinic can use. The materials tell you what to do at each stage, but not exactly how to tell AI to do it.
What's New
Last time, everything was provided: an analysis specification, verification targets, suggested prompts, a report template. You checked AI's output against expected values and caught a wrong denominator.
This time, the suggested prompts and the step-by-step specification are gone. You have a project plan that structures the work into stages, and verification targets that tell you what honest results look like. The rest is you directing AI through each stage.
The hard part is not the model. It is catching the moment when the model looks more accurate than it should. AI has a default that produces impressive-looking results on this kind of data. The verification targets will help you spot it.
Tools
- Python 3.11+ via your conda "ds" environment
- Jupyter Notebook for the analysis
- pandas for data handling
- scikit-learn for modeling and evaluation (new this project)
- matplotlib / seaborn for visualization
- Claude Code as the AI you direct
- Git / GitHub for version control
Materials
You receive:
- The extended dataset: 21 months of appointment records (~9,500 rows)
- A data dictionary describing every column
- A project plan that structures the prediction pipeline
- Verification targets for the prediction work
- A project governance file (CLAUDE.md) for Claude Code
Less than last time. No analysis specification, no suggested prompts, no report template. The project plan tells you the stages. The verification targets tell you what to check. You direct the work.