Capstone: an end-to-end project

Why a capstone?

Each module taught one skill. Real work needs all of them in the right order, and the hard parts often sit between modules: deciding which data is good enough, choosing a metric the business accepts, explaining a model's limits. The capstone is your chance to practise those joins, and it becomes the centrepiece of your portfolio.

Pick a good question

• Useful: someone would act on the answer.
• Answerable: data exists, or you can collect it ethically in a few weeks.
• Small enough: finishable in 3–4 weeks at an hour a day.

Example ideas: predict which food-delivery orders will be late (classification), forecast weekly vegetable prices in a mandi (regression or time series), classify customer support tickets into teams (NLP), or build a RAG assistant over your college's rules and timetables (generative AI).

The steps, with the module that covers each

Plan your time

At one hour a day, a sensible split for a four-week capstone is: week 1 for framing, data and cleaning; week 2 for exploration, the split and the baseline; week 3 for models and evaluation; week 4 for responsible-AI checks, the write-up and a demo. Write notes every day as you go. Most learners run out of time at the end, so start the README in week 1 and fill it in as each step finishes.

Presenting results

Lead with the answer, not the method: "The model flags 62% of late orders an hour early, at 80% precision, against 41% for the simple rule." Then show how you know (evaluation), what could go wrong (limitations, fairness, data gaps) and what you would do next. A non-technical reader should understand the first paragraph.

Summary
- Question: can we flag late deliveries 30 minutes before the promised time?
- Result: recall 0.62 at precision 0.80 on the last month (baseline rule: 0.41).
- Biggest signals: restaurant prep time, rain, rider distance.
- Limits: 3 months of data from one city; festival days under-represented.
- Next: add more cities, retrain monthly, monitor drift.

Common mistakes

• Choosing a question so big it never finishes. Shrink the scope.
• Spending all the time on models and none on data and evaluation.
• Ending with code only. Always write the summary.
• Skipping the responsible-AI checks because "it is just a project".