Choosing Portfolio Projects

Imagine you have finished a handful of AI experiments and want to put them in a portfolio. The natural instinct is to include all of them, since more work seems like more evidence. A portfolio, however, is not a complete archive of your activity. It is a selection meant to make a case about what you can do, and the reviewer reading it has limited time and limited attention. So the practical question becomes: how many projects should that selection contain?

The recommended number of projects for a portfolio is three. The reason given is that three sits between two competing pressures, and each pressure pushes in a different direction.

The first pressure favors fewer projects. Each project has to be strong enough that you can defend it in depth: explaining the problem, the choices you made, the trade-offs, and what you would change. Every additional project divides your attention and your ability to discuss the work fully.

The second pressure favors more projects. A single project, or only two, may not be large enough to demonstrate range, which means the spread of skills and kinds of problems you can handle.

Three is the number that satisfies both pressures at once. It is small enough that a reviewer can examine each project closely, and large enough that the collection can show more than one facet of your ability. The number itself is not the goal. It is useful because of what it allows.

Three-project guidance is built around two criteria. Judging any proposed portfolio comes down to asking whether each project can be defended in depth and whether the set as a whole shows range. Both questions must be answered, because a portfolio can pass one and fail the other.

The following example is an original illustration, not taken from the source. Suppose you have six finished AI experiments and want to choose a portfolio. Trace how the selection changes as you apply the two criteria.

When you are handed a proposed portfolio, or are reviewing your own, the method is the same. For each project, ask whether it contributes depth: can it be defended in its own right? For the set, ask what range it contributes and where projects overlap or leave gaps. The map below shows a proposed set of four projects evaluated this way. The projects and their relationships are an original illustration.

Read the map against the two criteria. The image classifier is the project you can defend in depth, and it contributes depth to the set. The classifier variant overlaps with it, since both address the same kind of problem, so it adds little range and divides attention without adding much evidence. The text summarizer and the data cleaning script each contribute range. The set as a whole therefore has the spread the guidance asks for, but the overlap means a reviewer would be looking at two projects where one would make the same point.

The three-project recommendation is not a rule about counting. It is a way of balancing depth on each project against range across the set. Once you can judge each project and the collection together, the number follows from the decision rather than the other way around.